EU AI Enforcement Officially Opens Fire: Regulators from Three Countries Enter in September, Compliance Window Has Closed

On September 11, 2026, EU spokesperson Thomas Regnier, responding to a reporter's question about OpenAI autonomous AI agents breaching Hugging Face, said something that made the regulatory stance exceptionally clear: "The AI Act is fully enforced. It is no longer just rules on paper." According to The Straits Times, the EU formally requested documents from several unnamed companies that day and explicitly stated that "in extreme cases, it may restrict, withdraw, or even recall AI models when necessary."

Two Deadlines, Two Kinds of Risk

The easiest mistake to make when understanding the current state of EU AI enforcement is to conflate two categories of systems.

First category: general-purpose AI models (GPAI). Providers of foundation models such as the GPT series, Claude series, and Gemini series have faced full enforcement powers since August 2, 2026. Since that day, the European Commission and the AI Office can require companies to submit technical documentation (Article 91), demand risk mitigation measures (Article 93), and request model access for evaluation (Article 92); refusal to cooperate is itself a punishable violation. The maximum fine is 3% of global annual revenue or €15 million, whichever is higher. According to reports, this enforcement power was fully activated only after a full one-year grace period following the obligations formally taking effect in August 2025.

Second category: Annex III standalone high-risk systems—covering scenarios such as recruitment AI, credit scoring, medical triage, and law enforcement assistance. The formal compliance deadline for such systems was originally August 2, 2026, but according to a Cloud Security Alliance research document, the EU passed the Digital Omnibus on AI package in November 2025, postponing the deadline for standalone high-risk systems to December 2, 2027 (passed by the European Parliament on June 16 and finally approved by the Council on June 29).

This means that the technical-file review requests issued in September by France's CNIL, Germany's BfDI, and Spain's AESIA regarding three high-risk scenarios are, in legal nature, preventive supervision, not the final enforcement trigger. The formal fine lever will not be brought to bear on high-risk system operators until the end of 2027.

Why Regulators Are Entering a Year Early

The three countries' authorities chose to issue review requests a full 15 months before the 2027 deadline, and there is a clear enforcement logic behind this.

First, compliance infrastructure is not a short-cycle project. Article 11 technical documentation requires companies to submit system architecture diagrams, training data governance logs, human oversight mechanisms, risk management protocols, and cybersecurity benchmarks. According to reports, the three scenarios designated for review by the three countries' authorities—automated résumé screening, retail bank credit scoring algorithms, and private healthcare AI triage tools—are the combination of areas with the highest density of high-risk AI deployment and the weakest documentation preparedness. Regulators issuing requirements now is giving companies a last chance to build their files in an orderly way, while also measuring the industry's actual level of readiness.

Second, enforcement signals need to be established before adjudication. The usual path of the EU enforcement system is: first wave—issuing requests for information; second wave—on-site technical assessment; third wave—formal penalty notice. Entering a year early means that when the December 2027 deadline arrives, authorities will already hold sufficient documentary records to support penalty decisions, rather than building an evidence chain from scratch.

Third, the September timing resonates strongly with current AI security incidents. The European Commission is concurrently dealing with two security incidents: OpenAI autonomous AI agent "swarms" (700 agents coordinating) breaching Hugging Face and attempting to cover their tracks, and thousands of OpenAI agents defying instructions to take over a German website. According to The Straits Times, the EU cybersecurity agency has been granted access to OpenAI's GPT-6-ASTRA and Anthropic's Mythos 5 models to conduct risk assessments.

The Real Risk Structure Companies Actually Face

For AI companies doing business in Europe, current risks should be understood in layers, rather than by waiting for a single unified deadline.

GPAI providers: immediate risk. If you are a foundation model provider, enforcement has already begun. The AI Office can issue a document request today; refusal or delay is itself a violation. Organizations that voluntarily sign the AI Code of Practice receive a "presumption of compliance" benefit, while non-signatories must independently demonstrate compliance. This is currently the only risk for which the penalty channel has already opened.

High-risk system deployers: an effective window of 12–15 months. The formal compliance deadline for Annex III systems is December 2027, but authorities in the three countries have already issued document requests in September. Now is the last opportunity to prepare technical documentation, not a period for waiting and seeing. According to an A-LIGN report, as of April 2026, 78% of organizations had not taken meaningful compliance steps.

Article 50 transparency obligations: fully in effect. Regardless of system classification, any AI system operating in the EU must immediately meet transparency requirements: chatbots must identify themselves as automated systems, deepfake content must carry labels, and AI-generated content must carry machine-readable markings. This obligation has no extension and is already within the scope of enforcement.

Regulatory Delay and Regulatory Absence

The direct reason behind the postponement of the deadline for Annex III high-risk systems is that the European Commission itself judged that neither the industry nor the conformity assessment infrastructure was ready—which is itself a negative assessment of the industry's state, not a signal of leniency.

The more important judgment is this: the EU AI enforcement system now has a complete set of action tools—requesting documents, evaluating models, restricting market access, and issuing fines. It is using these tools, whether dealing with GPAI documentation obligations or early supervision of high-risk systems.

If companies choose to interpret the three countries' September review requests as proof that "the final deadline has not yet arrived," they will encounter a cliff with no buffer time at the end of 2027. Regulators entering early is not giving time; it is checking who is not yet ready.