The criticism does not consist in denying the security safeguards of the French system. It asks what strategic value a standardized corpus of invoices could acquire if a barrier eventually gave way, and how to reduce that residual value.
METHODOLOGICAL WARNING This report systematically distinguishes three levels: facts documented by official texts, court decisions, public authorities or identified reports; plausible risk scenarios based on the known properties of data and systems; and forward-looking hypotheses, in particular the thought experiment of transposing the French system to the United States. It does not claim that any foreign State has compromised, will compromise or is currently seeking to compromise French electronic invoicing platforms. The issue is different: to determine the strategic value a corpus of invoices would have if it were one day exfiltrated, and to check whether the architecture was designed to limit that residual value. |
|---|
| “A serious security architecture must not only ask: how do we prevent intrusion? It must also ask: what will remain to be protected once one of the barriers has given way?” |
|---|
Executive summary
France’s generalization of electronic invoicing is neither a simple conversion from PDF to a structured format nor the creation of a single database in which the State would necessarily store every invoice in full. The system relies on an ecosystem of approved, specialized, interoperable platforms subject to security requirements. These platforms are meant to play a central role in issuing, transmitting and receiving invoices, as well as in transmitting certain invoicing, transaction and payment data to the administration. The texts provide in particular for ISO/IEC 27001 certification, obligations to operate within the European Union, restrictions on data transfers outside the Union and, where cloud hosting is used, requirements linked to SecNumCloud (the French qualification for trusted cloud services). These protections are substantial and must be acknowledged. [S1,S2,S3,S4]
The strategic criticism therefore does not claim that France designed a system without security. It concerns another level: when a country makes mandatory a standardized digital ecosystem covering the commercial relationships of millions of players, it mechanically creates points of concentration, interfaces, directories, histories, metadata and corpora likely to acquire a value far greater than that of each invoice taken in isolation. Yet a business invoice can identify the parties, the goods or services, quantities, prices and, depending on the case, other information that makes it possible to gradually reconstruct an economic graph. Among its official objectives, the reform also aims to improve knowledge of business activity. [S1,S5,S6,S36]
Artificial intelligence radically changes the risk calculation. A million stolen documents used to be a loot that was hard to exploit; modern systems can now extract entities, match suppliers, detect variations, rebuild dependency chains, identify weak signals and cross-reference a corpus with public information. A leak must therefore no longer be assessed only by the number of files stolen, but by the analytical power an adversary can apply to the whole. The working file behind this analysis already stressed this point: a structured invoicing database may interest anyone seeking to reconstruct customer-supplier graphs, dependencies, purchasing rhythms, prices, volumes or production ramp-ups; AI above all changes the economics of that exploitation. [S7,S8,S36]
China is a major case study because France, the United States and several independent bodies document at once the strategic importance Beijing attaches to AI, the existence of a sophisticated cyber ecosystem and operations attributed to actors reputed to be linked to Chinese interests. The French National Cybersecurity Agency (ANSSI) has thus described particularly dense activity by intrusion sets reputed to be Chinese for the purpose of capturing strategic and economic intelligence. The Cybersecurity and Infrastructure Security Agency (CISA) and several partner agencies published in 2025 a joint advisory describing actors backed by the People’s Republic of China targeting global telecommunications, government, transport and other infrastructure networks in order to support a global espionage system. These findings do not allow every attack to be blamed on Beijing or a future target to be presumed; they simply establish that the hypothesis of an interest in aggregated economic data is consistent with documented behavior. [S9,S15]
The most spectacular threshold crossed concerns cyber automation by AI. In November 2025 Anthropic published a report attributing with high confidence to a Chinese state-sponsored group a campaign in which Claude Code reportedly performed, according to the company, 80 to 90% of tactical operations. This attribution and these figures are Anthropic’s, not an independent judicial conclusion; they nonetheless constitute a major public case of cyber-espionage largely orchestrated by a software agent. In 2026, Reuters also reported that the Chinese company 360 presented Tulongfeng and Yitianzhen as national answers to advanced capabilities in vulnerability research and cyber defense, while Z.ai published results for GLM-5.3 close to Mythos 5 on certain vulnerability-discovery evaluations, but still clearly lower on exploitation. Performance announced by these companies must not be confused with independent validation. [S18,S20,S21]
China is, however, neither the only actor nor the only reason to secure the system. ANSSI also attributes espionage operations against French interests to Russian services, while cybercrime remains a daily threat. Private competitors, malicious insiders and foreign partners must also be taken into account. Economic warfare has precisely this peculiarity: diplomatic and military alliances do not dissolve commercial interests. In 2024, during a debate in the Senate, Roland Lescure summed up this principle bluntly: “in economic matters, there are neither friends nor allies.” A Senate report had already recalled that, even within the European Union, we are “allies but also competitors.” [S12,S13,S33]
The real French paradox then appears: the State officially recognizes the intensification of economic warfare, the need to protect industrial secrets, the risks of legal extraterritoriality, cyber-espionage campaigns and the vulnerability of small and medium-sized enterprises (SMEs); at the same time, it imposes the migration of a growing share of commercial relationships to a mandatory digital ecosystem. The answer must not be a simplistic rejection of all electronic invoicing. It must be a national doctrine of minimization, compartmentalization, encryption, limited exports, least-privilege control and an “assume breach” approach, that is, a design that assumes a barrier may one day be crossed and limits the consequences of a compromise in advance. [S4,S14,S35,S39,S40]
1. The Washington test: applying the French rule to American champions
To measure the scope of a national reform, it is often useful to take it out of its usual political environment. French electronic invoicing has become a technical subject, saturated with acronyms, timetables, formats, platforms and tax obligations. The debate naturally focuses on compliance: which platform to choose, which formats to accept, which data to transmit, which date to meet? This administrative approach is necessary, but it hides the question of sovereignty. The best way to make it visible is to mentally move the system to Washington.
Let us imagine that the US government announced to Tesla, SpaceX, Amazon, Anthropic and, eventually, to the whole American economy that business-to-business invoices could no longer circulate freely in their usual forms, but would have to go through an ecosystem of approved, interoperable intermediaries subject to obligations to transmit data to the authorities. Let us imagine that this rule came with perfectly legitimate objectives: fighting tax fraud, reducing administrative tasks, speeding up payments and improving the monitoring of economic activity. At first glance, the project could be presented as a modernization. Yet a question would immediately arise in boardrooms: what new intelligence surface does this architecture create?
Elon Musk would probably not reason in the same way for Tesla and for SpaceX, since the latter belongs to a world where advanced industry, space launchers, satellites, telecommunications and government contracts intersect. Jeff Bezos would examine the implications for Amazon, its supply chains, its suppliers, its cloud services and its immense ecosystem of partners. Dario Amodei would have to consider Anthropic’s commercial portfolio and the value that identifying large corporate customers, billing rhythms or variations in consumption would have for a competitor. The real purpose of the test is not to presume the personal reaction of these leaders; it is to make tangible the strategic nature of information that, when considered separately, appears purely accounting-related.
The American line of reasoning also allows a second experiment. Suppose a federal agency announced that all the necessary safeguards exist: security certification, audits, encryption standards, hosting requirements, access controls and penalties for breaches. Would the board of a sensitive company therefore accept the following conclusion: since the system is certified, there is no longer any need to ask what an adversary could do with the data after a compromise? Any mature cybersecurity culture would answer no. Certification reduces the probability and sometimes the impact of an incident; it never turns a complex system into an impregnable one.
| The right question is therefore not: “is an approved platform secure?” It is: “how much strategic intelligence can an adversary reconstruct if an account, an API, a contractor or a platform nevertheless ends up being compromised?” |
|---|
2. What France has actually built: neither PDF by email nor a single database
A credible criticism must begin by describing the system correctly. The French reform does not consist in automatically sending the State a full PDF copy of every invoice in a gigantic single national database. It relies on approved platforms that handle the issuing, transmission and receipt of electronic invoices, as well as the transmission of certain invoicing, transaction and payment data to the administration. Companies subject to the rules must use an approved platform for the flows concerned. As of September 1, 2026, all companies must be able to receive electronic invoices; large companies and mid-sized companies are required from that date to issue invoices and transmit the required data, while the issuing timetable for SMEs and micro-enterprises continues into 2027. [S1,S2,S3]
This architecture is therefore distributed, but standardized and interconnected. It is precisely this nuance that makes the debate interesting. A distributed ecosystem avoids some of the flaws of a single central database, but it does not abolish the risk of concentration: each platform retains or processes large sets of commercial relationships, the central directory and the interoperability mechanisms make routing easier, and the interfaces between systems become critical assets. A compromise would not necessarily have national scope; it could nevertheless produce a corpus large enough to be extremely useful to an adversary.
French texts also impose serious safeguards. The platform must in particular hold ISO/IEC 27001 certification covering the information system that supports its invoicing activity. It undertakes to operate this system from the territory of a European Union member State and to prevent the transfer of the service’s hosted data outside the Union. When a cloud provider is used, the regulations set requirements relating to the SecNumCloud qualification. Audits cover compliance, security, traceability and integrity. These mechanisms are not decorative; they constitute a real security foundation. [S4]
But acknowledging these protections does not close the subject. Cybersecurity has long relied on defense in depth. A company can have a firewall, multi-factor authentication, logs, a certification and competent teams while still being obliged to prepare for the scenario in which a legitimate account is stolen, an API (application programming interface) token is compromised or a supplier in the chain suffers an intrusion. The “assume breach” approach is not a judgment on the provider’s incompetence; it is an engineering method that refuses to confuse prevention with certainty. [S39]
This distinction must also be applied to the data themselves. The law sets out what must appear on the invoice, what must be structured and what must be transmitted to the administration. Article 242 nonies J provides that the invoicing data transmitted to the administration in structured form are chosen from among the mandatory mentions, with the exception of the precise description of the good or service. A change planned from September 1, 2027 nevertheless adds several pieces of information in structured format in the electronic invoice itself, notably the precise description, the quantity, the unit price and, depending on the case, the delivery address. The right reading is therefore neither “the State receives everything” nor “nothing commercial is structured.” It consists in mapping the flows precisely and applying a doctrine of minimal disclosure to everything that is not legally necessary. [S5,S6]
POINT OF RIGOR This report criticizes a risk of aggregation and compromise of a mandatory ecosystem. It does not claim that there is a single central database containing all the invoices of all companies in clear text. This clarification is essential: exaggerating the architecture would weaken precisely the economic security argument we are seeking to establish. |
|---|
3. The invoice as an economic sensor: what a corpus can reveal
A business invoice rarely contains a complete industrial secret. That is precisely why it is underestimated. Economic intelligence, however, often works by accumulation. A delivery address alone does not reveal an industrial strategy. An increase in quantity alone does not prove a ramp-up in production. The appearance of a new subcontractor alone does not demonstrate a change of technology. But when thousands of signals are aligned over time, the landscape changes.
A corpus of invoices can make it possible to reconstruct part of a company’s customer-supplier graph, to measure the concentration of activity on certain partners, to identify critical suppliers, to follow changes in prices or volumes, to observe seasonality, to detect the arrival of a new provider, the disappearance of an old one, an unusual increase in orders or a change of address. In certain sectors, carelessly detailed descriptions can reveal the names of projects, machines, prototypes, sites, products or operations. Attachments, comments and metadata can add even more context. This power of reconstruction is not purely theoretical: the OECD documents the use of VAT transaction data describing “who buys from whom” relationships to map buyer-supplier networks and production networks on a large scale. The French system is not identical to these foreign databases and does not necessarily transmit the same fields to the administration, but this example establishes the analytical value of a structured transactional corpus. [S7,S36]
There are therefore two levels of confidentiality. The first is documentary: what does a given invoice contain? The second is analytical: what can be deduced when documents are brought together? A cybersecurity policy that only protects the first level can fail at the second. A perfectly legal and ordinary invoice can become strategic information because it is correlated with a hundred thousand other invoices, with open data and with information from other leaks.
This logic explains why the real issue is not only the mandatory content. Order identifiers, project naming habits, attachments, free-text fields, PDF or XML metadata, contact addresses and the connections between invoicing, the CRM (customer relationship management), the ERP (enterprise resource planning software), the EDM (electronic document management) and the bank must also be examined. A technical connector with overly broad read rights can turn a localized intrusion into the invoicing system into a gateway to the company’s entire information assets.
The risk can be summed up by a simple notion: the marginal value of aggregation. A thousand invoices are not worth just a thousand times one invoice. They can be worth more, because the relationships between them are precisely what makes it possible to infer an economic structure. The phenomenon is comparable to certain mobility data: an isolated geographic position says little; a sufficiently long time series reveals habits, routes, sensitive places and relationships. [S36]
4. AI turns document theft into exploitable economic intelligence
The most important technological change may not concern how documents are stolen, but how they are exploited after the theft. Before the rise of large models and software agents, a massive exfiltration entailed a considerable human cost. Documents had to be indexed, companies recognized, formats harmonized, texts translated, terms searched, periods compared, graphs built and decisions made about which information deserved in-depth analysis. Volume could become an imperfect protection through overload.
This protection by volume is eroding. AI pipelines can convert heterogeneous documents into structured data, recognize entities, classify services, detect relationships, summarize changes, generate alerts on anomalies and cross-reference the results with open databases. The attacker no longer needs to read each invoice; they can ask the system to show the twenty suppliers whose spending has grown fastest, the customers whose billing frequency is collapsing, the new delivery sites, the companies linked to a sensitive sector or the projects whose names resemble a particular industrial nomenclature.
The point is not to claim that an AI automatically produces perfect truth. Models make mistakes, data can be ambiguous and correlation is not causation. But for intelligence purposes, an imperfect signal can be enough to steer a human investigation. AI therefore reduces the cost of pre-selection. It makes it possible to move from an ocean of documents to a list of targets, hypotheses and anomalies that analysts can then verify.
ANSSI already observes that espionage campaigns are a major component of the threat. In its 2024 overview, it stated that intrusion sets reputed to be linked to Chinese interests had been particularly dense and widespread for the purpose of capturing strategic and economic intelligence. CISA and its partners, for their part, described in 2025 actors backed by the People’s Republic of China seeking persistent access to global networks in order to feed an espionage system. These sources do not speak of French invoices; they only establish the context in which the value of an economic corpus must be assessed. [S9,S15]
| The problem is no longer only what an attacker can steal. The problem is what artificial intelligence allows them to understand, deduce, prioritize and exploit from what they have stolen. |
|---|
5. Tesla, SpaceX, Amazon, Anthropic: four scenarios to understand the value of the loot
Tesla: the supply chain as an industrial map
In an American scenario comparable to the French system, the interest of a corpus of Tesla invoices would not necessarily lie in discovering a technical blueprint. It could lie in gradually mapping suppliers, services, changes in volume and dependencies. A repeated increase in orders from a certain type of supplier, the arrival of an equipment maker, a change of destination or the multiplication of certain services can become signals when other open sources are available: job offers, permits, satellite images, logistics movements, supplier press releases, patents, technical publications and international trade data.
A competitor, a fund, an economic intelligence service or a State would not need immediate certainty. It would be enough to identify hypotheses: this supplier seems to be gaining importance; that plant seems to be increasing its output; this category of equipment appears; that expense is declining. Each hypothesis could then guide targeted collection. The invoice would thus become an instrument for reducing uncertainty.
SpaceX: when commercial intelligence touches strategic infrastructure
With SpaceX, the question becomes more sensitive because the company sits at the crossroads of commercial space, satellite communications, launchers and contractual relationships with American institutions. A corpus of invoices would obviously not give away the design secrets of Falcon, Starship or a satellite. But it could help identify specialized subcontractors, supply rhythms, the appearance of new services or certain changes in spending. In intelligence, discovering a discreet supplier can sometimes be more useful than reading a spectacular document, because it opens a new branch of collection.
The risk must therefore not be presented as “stealing the invoices would make it possible to rebuild a rocket.” Such a claim would be grotesque. The real risk is probabilistic: increasing an adversary’s visibility into industrial relationships, making it easier to select secondary targets and giving clues about changes in activity. In complex value chains, the most accessible link is not always the main company; it is often a smaller, less protected supplier that nonetheless holds a piece of the know-how or the schedule.
Amazon: logistics, cloud and bargaining power
Amazon is a third textbook case because the value of commercial data there is immense but different. Flows of suppliers, service providers, equipment, sites and spending rhythms can provide information on supply chains or infrastructure. For a company whose business model rests partly on large-scale optimization, knowing a competitor’s costs, dependencies or changes of provider better can become a commercial advantage.
Amazon Web Services (AWS) must not be forgotten either. A corpus of invoices relating to technology services can reveal business relationships and usage without exposing hosted content. In some markets, knowing who buys what, how often and at what approximate level of spending is already enough to improve a commercial strategy. A competing company can then focus its sales teams on identified customers instead of prospecting blindly.
Anthropic: the invoice as a competitive prospecting list
The case of Anthropic is particularly illuminating because the link between commercial data and competition is direct. Suppose an actor obtained a sufficiently large corpus of the company’s business-to-business (B2B) invoices. Depending on the actual structure of contracts and documents, it might be able to identify some of the corporate customers, observe billing rhythms, detect variations in usage or spending and distinguish the accounts that seem most strategic. It would not necessarily have all the customers, nor the full terms, nor an exact measure of the use of Claude. But it would already have reduced commercial uncertainty.
From there, a competitor could attempt targeted counter-offers: compute credits, temporary price cuts, migration assistance, preferential contract terms, integration services or personalized support. The stolen data would become an instrument for winning market share. The phenomenon is all the more interesting because the US FTC itself highlighted, in its study of partnerships between cloud providers and AI developers, the sensitivity of certain commercial information: finances, customer usage and revenue figures can affect competition when they become accessible to partners that also develop competing products. [S30,S31]
This analogy does not mean that the FTC equates an invoice with an antitrust violation, nor that illegal access to a corpus would automatically constitute a competition offense. It shows something simpler: the American regulator already regards access to non-public technical and commercial information as a variable likely to change competitive dynamics. The strategic value of an invoice database is therefore not a theoretical invention; it echoes a competition policy concern already documented in the AI economy. [S30,S31]
6. The FTC and competition: when access to information becomes a market issue
If Washington were considering a comparable national system, the Federal Trade Commission would obviously not have a magic power allowing it to cancel a federal law because it considered it dangerous. Its role would be different. The agency could contribute to the debate through its studies, its investigative powers, its expertise in competition and consumer protection, and by examining providers’ practices. The experience of its 6(b) inquiry into Microsoft-OpenAI, Amazon-Anthropic and Google-Anthropic shows precisely that it is interested in structures of dependency, switching costs, access to critical inputs and flows of sensitive information between partners. [S30,S31]
Transposed to electronic invoicing, the competition question would break down into several sub-questions. How many platforms concentrate the flows? Are there high exit or migration costs? Can a provider use data or metadata for secondary purposes? What barriers prevent abusive internal access? Can a customer’s information be used to profile its activity or to train a model? How would an incident affect the level playing field? What liability regime applies if trade secrets are exfiltrated and then exploited by a competitor? [S32]
The mere fact that the FTC has already requested detailed information on cloud-AI partnerships shows that, in the United States, competition authorities do not regard information flows as a peripheral issue. In its report, the agency notes that certain cloud partners may access sensitive information not available to others: models, development methods, chip co-design plans, financial data, customer usage and revenue. The general principle is transferable: the structure of access to information can create a competitive advantage, even without a legal merger between companies. [S30,S31]
This perspective deserves to be imported into the French debate. The security of a platform must not only be measured by the individual confidentiality of a document. It must also be assessed as market infrastructure: which players see which flows, for how long, with what export capabilities, what possibilities for aggregation and what mechanisms of proof in the event of an incident?
7. Economic warfare is not a metaphor: France says so itself
Describing these issues as “economic warfare” may seem excessive as long as the notion is confused with military warfare. Yet French institutions themselves use the term. A Senate report devoted to economic intelligence noted an increasingly fierce global competition, involving companies but also States, and recalled that influence games exist even within the European Union, where countries are both allies and competitors. [S12]
In July 2025, a National Assembly information report on economic warfare, focused on the defense industrial and technological base (BITD), described a confrontation using human, physical, digital, legal, capital and informational means. When presenting it, the rapporteur mentioned 500 to 550 proven attacks per year against the BITD and more than 750 economic security alerts in 2024, with strong exposure of SMEs. These figures do not describe the whole French economy: they concern a defense perimeter. But they illustrate the intensity of a threat that the State itself considers serious enough to measure and document. [S14]
Contemporary economic warfare therefore requires neither tanks nor missiles. It can take the form of a capital acquisition, targeted recruitment, a standard, a legal proceeding, an influence campaign, an export control, a cyber intrusion, a reputational attack or the capture of a secret. Its aim is not necessarily to destroy a company. It may be to make it dependent, delay its development, capture its market, acquire its know-how, weaken its financing or steer its decisions.
In this world, transactional data become a raw material. They provide information on what the company buys, sells, pays, subcontracts, accelerates or slows down. A nation that imposes the structured digitization of these flows must therefore regard them as a sovereign asset and not as a mere output of accounting software. [S36]
8. Diplomatic allies, economic competitors: the dark face of power relations
One of the hardest points to explain to the general public is that political alliance does not abolish economic competition. Two States can share a military alliance, cooperate on intelligence, defend common values and, at the same time, aggressively support their companies when they compete for a market. This does not mean that all competition is clandestine or illegal. It simply means that interests do not disappear behind declarations of friendship.
“In economic matters, there are neither friends nor allies.” Roland Lescure, French Senate, May 29, 2024. [S13] |
|---|
The sentence is particularly interesting because it does not come from an anti-American or anti-Chinese pamphlet, but from a French minister speaking before the Senate. In the same passage, he stressed that even allies and partners could use their security tools for economic intelligence purposes or to steer their investments. The Senate had already insisted on the sometimes harsh competition within the European Union itself. [S12,S13]
This observation forces us to abandon an overly simple representation of the world. China can be a strategic competitor and a trading partner. The United States can be a historic military ally and a major economic competitor. Germany, Italy or other European partners can cooperate within the Union while vigorously defending their industrial interests. India, the United Kingdom, Japan or South Korea can be security partners and rivals in certain markets. Economic geopolitics is a system of cross-cutting alliances, not a perfect separation between friends and enemies.
Questions of legal extraterritoriality illustrate this problem without any need to speak of illegal espionage. The French trusted cloud strategy and SecNumCloud include precisely criteria intended to limit the risk of access to sensitive data via non-European legislation. The Government has publicly defended these protections as a means of reducing exposure to foreign laws with extraterritorial reach. The message is clear: even when a provider belongs to an allied country, legal sovereignty and data protection remain issues in their own right. [S35]
The “dark face” of economic warfare is therefore not the idea that all allies are secretly enemies. Such a claim would be simplistic. It lies in the permanent coexistence of two planes: official cooperation and competition for money, influence, technologies, standards, supplies and market share. The citizen mainly sees the first. Sales departments, economic security services, diplomats and intelligence services also work on the second.
9. Why China deserves a specific analysis, without fantasy or naivety
China holds a special place in this report for three reasons. First, its industrial and technological mass gives it a structural interest in global value chains. Second, Beijing has made artificial intelligence an explicit strategic priority. Third, French and American authorities and several partners regularly attribute cyber-espionage campaigns to actors linked to the Chinese State. These three elements make it relevant to examine a scenario in which structured economic data would interest a Chinese actor. They do not justify turning every incident into an accusation against Beijing.
The Chinese “AI+” plan, published by the State Council in 2025, aims at a deep integration of AI into the economy, science, infrastructure, governance and society. The text highlights China’s advantages in terms of data, a complete industrial system and diversity of application scenarios, with milestones in 2027, 2030 and 2035. This plan is not an offensive document; it is a policy of economic and technological transformation. It nevertheless demonstrates the national scale at which Beijing thinks about AI. [S22]
MERICS, a European institute specializing in the study of contemporary China, analyzes this strategy as a quest for autonomy across the entire technology stack, from semiconductors to models and applications. Its 2025 report stresses that Beijing regards AI as strategic for national and economic security and seeks “independent and controllable” capabilities in a context of US export restrictions. This reading must be understood for what it is: the analysis of a European think tank, based on Chinese and open sources, not an official declaration of offensive intent. [S23]
There is also an official Chinese discourse in favor of responsible military AI. In its international positions, the Chinese Ministry of Foreign Affairs states that the weapon systems concerned must remain under human control, that humans must retain ultimate responsibility and that military AI must not be used to seek absolute supremacy or start wars. Including these texts is indispensable to avoid a one-sided demonstration: China does not publicly present itself as favoring an unregulated cyber race; on the contrary, it supports a governance framework and human control. [S24]
The analytical tension appears precisely between this official discourse, the documented strategic effort to develop technological autonomy and the operations attributed by other States to Chinese actors. International relations must be studied from these three levels at once: what a State says, what it builds and what its adversaries or partners say they observe.
10. From classic espionage to AI-assisted cyber-espionage
French authorities have already documented the economic dimension of cyber-espionage attributed to actors linked to China. ANSSI states in its 2024 overview that intrusion sets reputed to be Chinese carried out particularly dense and widespread activity for the purpose of capturing strategic and economic intelligence. The wording matters: the agency does not say that all Chinese operations pursue a commercial purpose, but it explicitly places economic intelligence among the objectives observed. [S9]
The United States describes a phenomenon of great magnitude. In 2025, CISA and several partner agencies published a joint advisory concerning actors backed by the People’s Republic of China targeting telecommunications, government, transport, hosting and military infrastructure networks around the world. According to this advisory, the actors seek long-term persistent access, notably in routers and network infrastructure, and the data obtained can allow Chinese intelligence services to identify and track the communications and movements of targets. [S15]
Two American court cases also show the economic dimension of technology capture. In January 2026, a federal jury found former Google engineer Linwei Ding guilty on several counts of economic espionage and theft of trade secrets related to artificial intelligence. The US Department of Justice states that more than two thousand pages of confidential information relating in particular to AI computing infrastructure were stolen, and that the jury found an intent to benefit interests linked to the People’s Republic of China. This is a verdict, not a mere intelligence attribution. [S16]
In March 2026, another case led to a $50 million fine against Hytera Communications after a guilty plea for conspiracy to steal proprietary technology from Motorola Solutions. The US Department of Justice explains that employees recruited from Motorola had taken information, including source code, allowing Hytera to develop products for a fraction of the cost incurred by Motorola. This example illustrates almost didactically what economic warfare is at company level: capturing know-how can cut years of research, development costs and time to market. [S17]
These cases obviously do not show that the Chinese State orchestrates every theft involving a Chinese national or company. They show that there are judicially established cases in which technological secrets were captured for the benefit of Chinese interests, and that Western authorities regard economic espionage as a national security risk. In this context, the potential value of a structured invoicing corpus cannot be assessed like that of a simple accounting file.
11. Mythos, Tulongfeng, GLM-5.3: an accelerated cyber race
The Anthropic case: when an AI agent orchestrates a large part of a campaign
In November 2025, Anthropic published a report it presents as the first documented case of a cyber-espionage campaign largely orchestrated by artificial intelligence. The company attributes the operation with high confidence to a Chinese state-sponsored group, designated GTG-1002. About thirty organizations were reportedly targeted and several intrusions reportedly succeeded. According to Anthropic, Claude Code was led to take part in reconnaissance, vulnerability discovery and exploitation, credential harvesting, lateral movement, analysis and exfiltration. The company estimates that the AI performed 80 to 90% of tactical operations, with human operators retaining certain critical decision points. [S18]
Two precautions are indispensable. First, the attribution to a Chinese state-sponsored actor is Anthropic’s; it is not a court judgment or a unanimously established international attribution. Second, the autonomy described does not mean that the system freely designed a geopolitical campaign. Humans had an objective, provided a framework and intervened at certain points. The interest of the case lies elsewhere: it shows that the human cost of an operation can be reduced when agents themselves chain together phases previously carried out manually.
This development is particularly important for electronic invoicing. Even if the intrusion itself remained classic, an agent capable of automating post-exploitation tasks could then index the documents, select commercial targets, produce a map of dependencies and generate intelligence notes. The attack and the analysis of the loot can thus be brought together in a single agentic chain.
360 Security: Tulongfeng and Yitianzhen
In June 2026, Reuters reported that the Chinese company 360 Security Technology presented two tools grouped under the name “Yitian Tulong.” Tulongfeng is described as a system for automatic vulnerability discovery, while Yitianzhen focuses on cyber defense and incident response. The founder of 360, Zhou Hongyi, explicitly presented this capability as strategic and mentioned the risk of “one-way transparency” if American systems could analyze Chinese infrastructure without a national equivalent. [S20]
360 claims that Tulongfeng discovered several thousand vulnerabilities, more than a hundred of which were reportedly verified by the Chinese authorities. Reuters said it had not been able to confirm these results independently. One must therefore resist the temptation to turn a company announcement into proven technological equivalence. The reasonable conclusion is more limited but already significant: a major Chinese player is explicitly working to produce agentic tools specialized in vulnerability research and presents this skill within a framework of international strategic competition.
Z.ai and GLM-5.3: close on discovery, still far behind on exploitation
In August 2026, Z.ai published results for GLM-5.3 on several cyber evaluations. Reuters reported that the model scored 84.5% on CyberGym, against 83.8% for Mythos 5 according to the figures compared by the company, but only 54.4% on ExploitBench against 78% for Mythos 5. The difference is essential: detecting or understanding a vulnerability and producing a working exploit are two different levels of capability. [S21]
Here again, the results come from the developer and have not undergone full independent verification. They nevertheless show that the competition can no longer be summed up as “the United States has the advanced models, China follows far behind.” According to MERICS, the Chinese ecosystem is closing the gap on certain segments of models and applications despite constraints on advanced chips. Specialized systems such as those of 360 and more general models such as GLM-5.3 point to a diversification of the Chinese approach. [S21,S23]
For a State that wants to protect its economic assets, the conclusion must not be to engage in alarmist one-upmanship over the exact level of each model. Benchmarks will change. Names will change. Performance will improve. The timeless question is this: should an infrastructure designed to last ten or twenty years be secured according to the offensive capabilities of 2026, or according to the reasonable assumption that cyber automation and document analysis capabilities will progress strongly throughout its lifetime?
12. Who really observes China? Institutions, foundations, think tanks and private intelligence
The partial opacity of the Chinese system makes research difficult. Many intelligence and cyber defense capabilities are, by nature, not public. It would therefore be dangerous to rely on a single source, especially when it belongs to a country engaged in strategic competition with Beijing. The most robust method is to cross-reference several families of sources: official Chinese documents, French authorities, American and allied agencies, court decisions, academic institutes, specialized think tanks, cyber intelligence companies and the international press.
| Organization | Nature | Main contribution | Reading caution |
|---|---|---|---|
| ANSSI | French public authority | Annual overviews, incident response, technical attributions and recommendations. Useful for the targeting of France and for distinguishing cybercrime from state espionage. | French state source; does not cover all classified intelligence. |
| DGSI | French domestic intelligence service | “Flash ingérence” alerts, economic security, raising companies’ awareness of methods used to capture know-how and information. | Many cases deliberately anonymized; priority given to prevention. |
| CISA and partner agencies | US and allied cyber authorities | Joint technical advisories on intrusion sets, compromised infrastructure and state-sponsored actors. | Government attributions; to be cross-checked with other sources. |
| DOJ / FBI | US justice and counter-intelligence | Indictments, guilty pleas, verdicts and economic espionage cases. | Very high factual value when a judgment or plea exists; US institutional communication. |
| MERICS | European think tank specializing in China | Industrial policy, AI, technological autonomy, economy, governance and security. | European analysis; useful for contextualizing Chinese texts. |
| CSET, Georgetown University | US academic research center | Analysis of public procurement and open documents concerning AI, semiconductors and Chinese military needs. | Open-source work; some source documents disappear or are incomplete. |
| U.S.-China Economic and Security Review Commission | US Congressional commission | Hearings and reports on data, technology, economy, cybersecurity and the Sino-American balance of power. | An openly American strategic perspective; testimony must be distinguished from judicial facts. |
| ASPI | Australian think tank | Economic cyber-espionage, critical technologies, regional security and influence. | Security think tank; analysis to be cross-checked with official and academic sources. |
| Jamestown Foundation | US foundation / think tank | China Brief, monitoring of the PLA, the CCP, military AI and security systems. | Open sources and specialized analyses; some conclusions are interpretive. |
| Recorded Future / Insikt Group | Private threat intelligence | Cyber infrastructure, attributed groups, capability providers and observed links between PLA/MSS ecosystems and contractors. | Commercial source; useful for technical details but attribution must be corroborated. |
| Anthropic Threat Intelligence | Private AI company | Model abuse, agentic campaigns and attack techniques using Claude. | Privileged view of its own systems; attribution and metrics remain the company’s. |
| Reuters | International news agency | Journalistic verification of announcements by Chinese providers and explicit reporting of results not independently verified. | Does not replace a technical audit, but provides important independent mediation. |
| CAC / State Council / Chinese MFA | Official Chinese sources | AI+ plans, governance doctrine, data security, positions on military AI. | Indispensable primary sources for knowing Beijing’s public position; obviously they do not describe clandestine operations. |
None of these sources should be treated as absolute. An American agency works within the framework of American interests; a think tank has its institutional culture; a cybersecurity company sells services; a government sets out its public position; an AI company describes the abuses it observes on its own product. Robustness comes from convergence. When French, American, Chinese and independent sources all describe, from different angles, an acceleration of AI, a growing strategic value of data and an intensification of technological competition, public decision-makers can no longer treat these dimensions as an optional add-on to tax reform.
CSET has thus exploited thousands of requests for proposals and contracts linked to the Chinese People’s Liberation Army (PLA) to document AI needs in command, intelligence, surveillance, cyber and other functions. Another study by the same center analyzes the mobilization of civilian actors for military AI needs. The Jamestown Foundation, for its part, has catalogued procurement documents and analyses linked to the integration of systems such as DeepSeek into certain military or public security uses. The nuances must be kept: this work shows adoption dynamics and ambitions; it does not prove that every Chinese AI company works for the PLA or that every civilian model is a military tool. [S25,S26,S27]
In April 2026 the U.S.-China Economic and Security Review Commission devoted an entire hearing to China’s data strategy, under the title “Taking a Bigger Byte: China’s Expanding Strategy for Data Dominance.” The very subject of this hearing shows that, in the American strategic community, data are now thought of as an asset of power, not merely as a computing resource. This commission belongs to the US Congress and its outlook is naturally shaped by Sino-American rivalry; its work must therefore be used as one element of the file, not as a neutral arbiter. [S28]
Finally, Recorded Future and other threat intelligence companies have documented an ecosystem in which groups attributed to the Chinese Ministry of State Security or to the People’s Liberation Army use infrastructure, software, capability providers and sometimes private contractors. Here again, confidence levels vary from case to case. But this public-private hybridization is precisely one of the reasons why modern cyber attribution is complex: a State may have a range of actors more or less directly linked to its services. [S29]
13. China is not the only threat: Russia, cybercrime, insider threats and private competitors
Focusing the analysis on China must not create a blind spot. In July 2026, ANSSI made public an attribution concerning the Turla intrusion set, which it links to the 16th Center of the Russian FSB. The agency describes intelligence-gathering activities against government, diplomatic, defense, research, technology, education, media and energy sectors, and specifies that French strategic, diplomatic and economic interests are targeted. [S33]
Cybercrime also remains a massive threat. Ransomware and extortion through data theft seek financial gain first, but the result of a leak can then circulate, be resold, published or reused. A competitor does not need to have commissioned an attack itself to try to exploit data that have become accessible. Trade secret and unfair competition law can offer remedies, but legal remedies come, by definition, after a loss of confidentiality or an unlawful exploitation has been established.
The insider threat must also be taken into account. An employee, an administrator, a subcontractor or a provider with legitimate rights can become a source of exfiltration. Security must therefore not be designed only against an outside hacker. Least privilege, separation of roles, export limits, logging and dual validation of sensitive operations also protect against the abuse of an authorized account. [S39]
Finally, the ordinary economic competitor must remain in the threat model. Most competitors respect the law and obviously will not buy stolen data. French law protects as trade secrets information that is not generally known or readily accessible, that has commercial value because it is secret and that is subject to reasonable protection measures. Preparing proof of these protection measures is therefore part of economic cybersecurity. [S37]
14. French incidents recall a simple truth: certification does not abolish risk
In August 2026, the CNIL (the French data protection authority) announced that it had been notified of data breaches affecting the information system of the Direction générale des Finances publiques (DGFiP, the French public finance directorate). According to the authority, a third party had been able to consult and extract tax and land registry information relating to individuals and businesses, including certain income data, SIREN numbers and company names. This incident does not concern the electronic invoicing platforms and in no way allows the conclusion that they will suffer the same fate. It only recalls that a sensitive administration subject to security requirements can be the victim of a breach. [S10]
A few months earlier, the Ministry of the Interior had announced an incident on the ANTS portal (the national agency for secure documents) that could concern 11.7 million accounts. The investigations then ruled out certain additional data such as attachments and biometric data, but the episode once again illustrated the difficulty of guaranteeing the inviolability of a large exposed public system. [S11]
These examples must be used with restraint. A serious cybersecurity policy does not consist in saying: “since the DGFiP or the ANTS have had an incident, all approved platforms will be hacked.” It consists in saying: “since critical systems can be compromised despite protections, the design must treat compromise as a scenario, not as an impossibility.” That is exactly the philosophy of defense in depth.
Compliance with a framework must also be distinguished from technical invulnerability. The fact that a platform is ISO/IEC 27001 certified for the required scope is not equivalent to a guarantee that it has no vulnerabilities. ANSSI itself specifies that SecNumCloud addresses specific threats and that a digital service hosted on a qualified cloud offering does not automatically inherit that qualification. The SecNumCloud framework further states that, to protect technically against certain accesses by the provider to the data, additional encryption means under the customer’s control may be necessary. [S4,S40]
15. The Italian precedent: minimizing what the administration retains and exploits
Italy generalized electronic invoicing before France and its experience is instructive. The Garante per la protezione dei dati personali, the Italian data protection authority, criticized certain initial aspects of the system, notably the generalized archiving of invoice content. Subsequent changes led to reflection on limiting storage to the tax data needed for certain audits and on protecting sensitive descriptive fields. [S34]
The Italian precedent cannot be transposed word for word into French law. Architectures, texts, obligations and timetables differ. But it establishes a fundamental design principle: the commercial recipient may need information that the administration or an intermediary does not need to retain or exploit in clear text for all purposes. This separation between content needed for the contract and data needed for tax auditing is a particularly important privacy-by-design avenue for the future.
It ties in with selective disclosure and confidential computing technologies: end-to-end encryption of certain fields, confidential computing, secure multi-party computation, zero-knowledge proofs or homomorphic encryption for certain checks. Not all these techniques are ready to replace the French architecture tomorrow, and some remain costly or complex. But the fact that they exist rules out regarding clear-text visibility of all data as an eternal technological inevitability.
16. The questions France should have asked before September 1, 2026
The timetable being what it is, the point of a timeless article is not to replay the debate as if September 1, 2026 had never happened. It is to formulate the questions that should have structured the design before that date and that remain valid after it. A national infrastructure does not become frozen on the day it comes into force; its standards, audits, contracts, technical requirements and practices can evolve.
What minimum amount of commercial information really needs to remain accessible in clear text to each intermediary, and for how long?
Can a compromised user or administrator account export tens or hundreds of thousands of invoices without immediately triggering a crisis?
Are the rights to issue, read, export and administer technically separated, or can a single account accumulate excessive privileges?
Are encryption keys separated by customer, by logical customer environment (“tenant”), by period or by sensitive portfolio, so as to reduce the blast radius of a compromise?
Are access and export logs replicated outside the platform so that an attacker cannot erase the evidence at the same time as the data?
Are the invoicing APIs designed according to least privilege, or can they siphon off the entire CRM, contracts, EDM or ERP?
Is there a national notification and transparency mechanism that makes it possible to quickly know the volume, nature and economic scope of a leak?
Can the platforms use the data for secondary purposes, profiling, training or commercial analysis, and do the contracts explicitly prohibit this where it is not necessary?
Does the system require regular tests by a “red team” reproducing the realistic scenario of a stolen legitimate account, rather than only checking external vulnerabilities?
Do companies have an exit and revocation procedure fast enough to change platform when a major incident occurs?
Is there a doctrine of liability and compensation suited to trade secret damage, which can be far greater than the nominal value of an invoice?
Does France sufficiently fund research on selective disclosure, confidential computing and cryptographic techniques allowing audits without unnecessary exposure of data?
Can sensitive companies compartmentalize their flows by establishment, portfolio or activity without creating complexity that further increases the attack surface?
Do the regulations distinguish clearly enough between the data indispensable for tax audits, those needed by the commercial recipient and those that should never enter the invoicing channel?
These questions are not a plea for standing still. The digital transformation is already under way and, if well designed, can improve security compared with heterogeneous practices in which unprotected PDFs circulate by email. But this argument is only valid if functional centralization and standardization do not create an even greater economy of scale for the attacker. The right objective is not a return to email; it is an architecture in which the structured invoice is less exploitable en masse than a set of poorly secured commercial files.
17. What companies can still do: reduce the strategic value of a leak
The mandatory nature of the system does not condemn companies to passivity. A large part of the risk comes from information they add themselves beyond what is necessary, from the way they name their projects and from the breadth of technical access. The realistic strategy does not consist in making a leak harmless; that is impossible when the identity of the parties, prices and other mentions must legally appear. It consists in preventing an invoice from becoming a complete commercial intelligence file.
17.1. An invoice must remain an invoice, not a project file
A minimalist invoicing doctrine must remove free comments, internal prototype names, technical details, employee names, operational explanations and attachments that are not legally or contractually necessary. The contract, the service report, the plan, the bill of materials and the minutes must circulate through a separate, authenticated and compartmentalized channel. [S7]
The principle can be summed up as a two-layer architecture. The first contains the legal skeleton of the transaction: parties, operation, price, VAT, payment and required mentions. The second contains commercial and technical knowledge: purpose of the project, execution details, plans, staff, specifications, results, schedule and secrets. A neutral identifier links the two without turning the invoice into a semantic index of the secret.
17.2. References must identify without telling a story
A code like “AIRBUS-DEFENSE-RADAR-2027” or “PROJET-NUCLEAIRE-FLAMANVILLE” already says too much. A non-semantic reference, generated and managed in a separate repository, reduces the value of a stolen corpus. The correspondence table then becomes the real vault: when the stakes are high, it must not be stored in the same tenant, the same cloud, the same account or the same administrative environment as the invoicing.
17.3. APIs must never become vacuum cleaners
The invoicing chain is often connected to the CRM, the ERP, the bank, document management and accounting tools. Each integration increases productivity but can also increase the blast radius of a compromise. A technical account intended to issue an invoice should not have full read rights over the CRM. An issuing token should not be able to export ten years of history. API secrets must be separated by function and be revocable immediately. [S7,S39]
17.4. Prohibit mass export as ordinary behavior
One of the simplest protections to demand consists in making abnormal what is abnormal. An accountant does not need, in their daily work, to download two hundred and fifty thousand invoices in ten minutes. Platforms should allow quotas, temporary privilege elevation, a second approval, a security delay and an out-of-band alert for large exports. A platform unable to detect mass behavior should not be judged only on its ability to protect the password.
17.5. Build the proof of trade secrecy now
For information to benefit from trade secret protection, French law requires in particular that it be subject to reasonable protection measures, given the circumstances. A simple “confidential” label is therefore not enough, on its own, to demonstrate the existence of such measures. Sensitive categories must be defined, access controlled, the need to know documented, systems segmented, consultations logged, contractual clauses provided for and authorizations managed. Cybersecurity then also prepares the legal proof that the information was effectively protected. [S37]
17.6. Think in terms of hot storage and cold storage
The French Commercial Code requires accounting documents and supporting documents to be kept for ten years. This retention period does not mean that ten years of invoices must remain instantly searchable from every application. Where the architecture, the contract and the applicable obligations allow it, a historical archive can be separated from operational storage, subject to strengthened authentication and stripped of ordinary mass search. The aim is that a compromise of the everyday system should not automatically give access to the company’s entire history. [S38]
17.7. Test the circuit breaker before you need it
A crisis procedure must make it possible to revoke all sessions, disable API tokens, suspend sensitive accounts, cut connectors, renew authentication secrets and contact the platform through a channel independent of the compromised account. The test must be carried out in real conditions. An emergency procedure that has never been executed is often a hypothesis, not a capability.
17.8. Demand continuous improvement from the State and the platforms
At national level, several improvements deserve study: a confidential invoice profile, a clearer separation between the tax envelope and the commercial envelope, end-to-end encryption of fields not needed by the intermediary, separate keys per company or period, a ban on global exports by a single account, quotas and exfiltration detection, regular “red team” exercises, publication of incident indicators, mandatory forensic capability, compensation mechanisms and research on privacy technologies. The technical file that preceded this article already set out this doctrine: protect the invoice, the commercial graph, identity, interfaces and the archive, then prepare the legal response. [S7,S39,S40]
18. Conclusion: if Washington would hesitate, why shouldn’t Paris have doubts?
French electronic invoicing must not be judged through a caricatured opposition between modernity and the past. Email and PDF are not models of security. VAT fraud is a real problem. Dematerialization can reduce costs, improve accounting processing and speed up exchanges. Approved platforms are subject to real security requirements. None of these elements should be denied. [S1,S4]
But the debate becomes incomplete as soon as it stops there. France operates in a world where its own authorities speak of economic warfare, where ANSSI documents strategic and economic espionage, where the DGSI raises companies’ awareness of foreign interference, where foreign services target digital infrastructure, where AI secrets are the subject of thefts and court proceedings, and where the best artificial intelligence systems are beginning to automate a substantial part of cyber operations and document analysis. [S8,S9,S16,S18,S19,S33]
In this world, the question France should have asked itself before September 1, 2026 was not only: “how do we make invoices electronic?” It was: “how do we make an economic data ecosystem mandatory without creating, for the day one of its links is compromised, a disproportionate advantage for an adversary, a competitor or an intelligence service?”
The Washington test retains all its force here. If the American administration announced a comparable system, Tesla, SpaceX, Amazon and Anthropic would probably demand guarantees on access, retention, compartmentalization, liability, extraterritoriality and an attacker’s ability to aggregate the data. The FTC would probably take an interest, within its remit, in concentration effects and in the risks that access to sensitive information poses to competition. Security agencies would examine the consequences of a compromise by a state actor. Why would these questions be excessive when they concern French companies?
China is a potential adversary in some scenarios, a partner in others, a major technological competitor and an actor whose cyber capabilities are constantly monitored. Russia is another documented threat. Cybercriminals do not need a flag. A person with legitimate but malicious access, an “insider threat,” can open the door from the inside. And allied countries also continue to defend their economic interests. National security must therefore not be built against a country’s name; it must be built against classes of capabilities and intentions.
That is why the best doctrine is neither blind trust nor catastrophism. It fits in one sentence: design as if a barrier will one day give way, then make sure the forced door never opens onto the whole building.
The reform was designed to know economic activity better and make it flow more smoothly. It must now be assessed with the same ambition in the other direction: preventing an unauthorized actor from using the same flows to know the French economy better than France would ever wish to allow. [S1]
| The question is not whether France had to modernize. The question is whether it has given enough thought to the strategic value of what it decided to standardize, interconnect and make mandatory. |
|---|
Appendix A. Map of reference organizations
The following table classifies the main sources used to follow China, the cyber threat and economic warfare. It does not claim to be exhaustive; its purpose is to give the reader a diversified monitoring method.
| Country / type | Organization | Field covered | Recommended reading |
|---|---|---|---|
| France | ANSSI | Cyber threat, attributions, incidents, recommendations | Very high for French technical facts |
| France | DGSI | Economic interference, counter-espionage, raising companies’ awareness | High; cases often anonymized |
| United States | CISA | Joint cyber advisories, infrastructure and intrusion sets | High for an allied government position |
| United States | DOJ / FBI | Court cases on trade secret theft and economic espionage | Very high when there is a verdict/plea |
| United States | FTC | Competition, sensitive data, cloud-AI partnerships | Very high within its regulatory remit |
| Europe | MERICS | Chinese economy, technology, AI and security | Specialized think tank; to be cross-checked |
| United States | CSET Georgetown | PLA procurement, military AI, semiconductors | Academic open-source research |
| United States | USCC | Sino-American economics and security, Congressional hearings | American strategic perspective |
| Australia | ASPI | Economic cyber-espionage, critical technologies | Security think tank; to be cross-checked |
| United States | Jamestown Foundation | CCP, PLA, security, China Brief | Specialized open-source analysis |
| Private | Recorded Future / Insikt | Technical intelligence on threats and infrastructure | Commercial; useful with corroboration |
| Private | Anthropic Threat Intelligence | Abuse of Claude and agentic campaigns | Proprietary data; its own attribution |
| China | CAC / State Council | Official economic and digital plans | Primary public policy source |
| China | Ministry of Foreign Affairs | Public position on AI, security and governance | Primary diplomatic source |
| International | Reuters | Journalistic verification of announcements and their limits | Independent secondary source |
Appendix B. Matrix: documented facts, plausible scenarios and limits
| Proposition | Assessment | Limit / caution |
|---|---|---|
| The French system relies on approved platforms and provides for data transmissions to the administration. | DOCUMENTED FACT | Official tax sources [S1,S2,S3]. Does not mean that a single database contains all invoices. |
| An invoice can reveal parties, price, quantity and other information depending on the context. | DOCUMENTED FACT | The exact content depends on the type of transaction and the mandatory mentions [S5,S6]. |
| Aggregating thousands of invoices can help reconstruct customer-supplier relationships and economic rhythms. | PLAUSIBLE ANALYTICAL SCENARIO | Accuracy depends on the corpus, the available fields and cross-referencing with other sources. OECD work shows that VAT transaction data can be used to map buyer-supplier networks [S36]. |
| A foreign service could be interested in an invoicing corpus if it obtained one. | RISK SCENARIO | No evidence is presented here that any service is currently targeting French platforms. |
| ANSSI documents intrusion sets reputed to be Chinese with a strategic and economic intelligence purpose. | DOCUMENTED FACT | French technical/government attribution, not attribution of every attack to Beijing [S9]. |
| Anthropic attributes to a Chinese state-sponsored group a campaign in which Claude Code reportedly performed 80 to 90% of tactical operations. | FACT REPORTED BY ANTHROPIC | Attribution and metrics are Anthropic’s own; not an independent judgment [S18]. |
| 360 claims to have Tulongfeng and Yitianzhen, advanced cyber tools. | COMPANY ANNOUNCEMENT | Reuters did not independently verify the announced performance [S20]. |
| GLM-5.3 approaches Mythos 5 on certain vulnerability-discovery benchmarks. | VENDOR RESULTS REPORTED BY REUTERS | Large gap on exploitation; results not independently validated [S21]. |
| Allies can be economic competitors. | POLITICAL AND ECONOMIC OBSERVATION | Does not mean that allies systematically conduct illegal operations [S12,S13]. |
| ISO 27001 or SecNumCloud certification makes a platform impregnable. | FALSE | These schemes reduce risk and impose safeguards; they do not promise the absence of any incident [S4,S40]. |
Sources and selected bibliography
The references below were preferred for their official, judicial, academic or explicitly attributed character. The report’s forward-looking analyses are flagged as such in the text. Reference S7 is an internal technical doctrine note by the author: it does not constitute independent corroboration and is now supplemented, on the essential points, by external legal, institutional and academic sources [S36 to S40]. Links consulted and checked: August 21, 2026.
[S1] DGFiP, “Je découvre la facturation électronique” (Discovering electronic invoicing): scope, objectives and role of the platforms. Consult the source
[S2] DGFiP, implementation timetable of the reform. Consult the source
[S3] DGFiP, “Facturation électronique et plateformes agréées” (Electronic invoicing and approved platforms): central role, e-invoicing and e-reporting. Consult the source
[S4] Légifrance, obligations applicable to approved platforms: ISO/IEC 27001, SecNumCloud, operation within the EU and transfer outside the EU. Consult the source
[S5] Légifrance, structured invoicing data transmitted to the administration, Article 242 nonies J and related provisions. Consult the source
[S6] Légifrance, additional structured data applicable from September 1, 2027. Consult the source
[S7] Author’s technical doctrine note: invoice minimization, “assume breach” approach, trade secrets, APIs, compartmentalization and countermeasures. Version of August 14, 2026, internal working document.
[S8] DGSI, Flash ingérence: economic interference actions targeting French companies. Consult the source
[S9] ANSSI, 2024 Cyber Threat Overview: cybercrime, Russia, China and the capture of strategic and economic intelligence. Consult the source
[S10] CNIL, “Piratage du système d’information des impôts : les vérifications sont en cours” (Hacking of the tax information system: checks under way), August 18, 2026. Consult the source
[S11] Ministry of the Interior, ANTS portal security incident, update of April 21, 2026. Consult the source
[S12] French Senate, “Anticiper, adapter, influencer : l’intelligence économique comme outil de reconquête de notre souveraineté” (Anticipate, adapt, influence: economic intelligence as a tool for regaining our sovereignty), 2023. Consult the source
[S13] French Senate, sitting of May 29, 2024: debate on economic intelligence and statement by Roland Lescure. Consult the source
[S14] National Assembly, information report on economic warfare, July 16, 2025, No. 1757. Consult the source
[S15] CISA and partners, “Countering Chinese State-Sponsored Actors Compromise of Networks Worldwide to Feed Global Espionage System,” 2025. Consult the source
[S16] U.S. Department of Justice, Linwei Ding verdict: economic espionage and theft of AI technology, January 30, 2026. Consult the source
[S17] U.S. Department of Justice, Hytera fined $50 million for conspiracy to steal Motorola technology, March 9, 2026. Consult the source
[S18] Anthropic, “Disrupting the first reported AI-orchestrated cyber espionage campaign,” full report, November 2025. Consult the source
[S19] Anthropic, threat intelligence work on the agentic automation of attacks, 2026. Consult the source
[S20] Reuters, “China’s 360 says it has developed tools to match Anthropic’s Mythos,” June 24, 2026. Consult the source
[S21] Reuters, “China’s Z.ai says new model nears Anthropic’s Mythos 5 in cyber-defence tests,” August 14, 2026. Consult the source
[S22] Chinese State Council / CAC, “AI+” plan, August 27, 2025. Consult the source
[S23] MERICS, “China’s drive toward self-reliance in artificial intelligence: from chips to large language models,” July 22, 2025. Consult the source
[S24] Chinese Ministry of Foreign Affairs, position on regulating military applications of AI and human control. Consult the source
[S25] CSET, Georgetown University, “China’s Military AI Wish List,” 2026. Consult the source
[S26] CSET, “Pulling Back the Curtain on China’s Military-Civil Fusion: How the PLA Mobilizes Civilian AI for Strategic Advantage,” 2025. Consult the source
[S27] Jamestown Foundation, “DeepSeek Use in PRC Military and Public Security Systems,” 2025. Consult the source
[S28] U.S.-China Economic and Security Review Commission, hearing “Taking a Bigger Byte: China’s Expanding Strategy for Data Dominance,” April 30, 2026. Consult the source
[S29] Recorded Future / Insikt Group, “Charting China’s Climb as a Leading Global Cyber Power,” 2023. Consult the source
[S30] Federal Trade Commission, “FTC Issues Staff Report on AI Partnerships & Investments Study,” January 17, 2025. Consult the source
[S31] FTC Office of Technology, “Behind the FTC’s 6(b) Report on Large AI Partnerships & Investments,” 2025. Consult the source
[S32] FTC, “AI Companies: Uphold Your Privacy and Confidentiality Commitments,” 2024. Consult the source
[S33] ANSSI, “Ciblage et compromission d’entités françaises par le Centre du FSB” (Targeting and compromise of French entities by the FSB Center), July 13, 2026. Consult the source
[S34] Garante per la protezione dei dati personali, FAQ “Fatturazione elettronica”: full storage, minimization, integrity and confidentiality. Consult the source
[S35] National Assembly / Government, answer on trusted cloud, SecNumCloud and protection against extraterritorial legislation. Consult the source
[S36] OECD, “Transaction data for evidence-based industrial policy” (2025): VAT transaction data as a tool for analyzing inter-company relationships and production networks. Consult the source
[S37] Légifrance, French Commercial Code, Article L151-1: criteria for trade secret protection, including reasonable protection measures. Consult the source
[S38] Légifrance, French Commercial Code, Article L123-22: ten-year retention of accounting documents and supporting documents. Consult the source
[S39] ANSSI, “Défense en profondeur” (Defense in depth), 2026: least privilege, access control, encryption, supply chain and regular auditing. Consult the source
[S40] ANSSI, SecNumCloud FAQ and framework: scope of the qualification, customer’s residual responsibilities and limits requiring additional measures depending on the case. Consult the source

