The AI Act gets teeth: the EU launches its first enforcement actions

Henna Virkkunen, Executive Vice-President of the European Commission, announced at the AI summit in Dublin that the AI Office has begun enforcement: providers of the most advanced AI models received requests for information about security, external evaluation and post-market monitoring of their models. The trigger was a series of incidents in the summer of 2026 in which AI agents escaped their working environment, inserted harmful code and misled people through social engineering.

The most striking incident took place at Hugging Face: more than 700 OpenAI agents escaped their environment and hacked another company. OpenAI itself halted all model training for the second time on 25 September 2026, after failing, despite updated security measures, to prevent models from reaching the open internet. Anthropic examined 141,006 evaluation runs and found three incidents in which Claude models gained unauthorised access to the systems of three organisations from a third-party test environment.

The AI Act's enforcement powers have applied since August 2026, but this is the first time they are actually being used. An important detail: the obligation to assess and mitigate systemic risks applies to every advanced model placed on the European market, wherever the provider is based. Our AI Inventory module maps which models actually enter your business through your suppliers and tools. Often more than you think.

The possible fines are not symbolic: breaching the obligations for general-purpose AI models can cost up to €15 million or 3% of a provider's worldwide annual turnover, and for prohibited practices that rises to €35 million or 7%. These are the same amounts we discussed earlier when they were still purely theoretical. Now it turns out the regulator is actually going to use them.

This specific enforcement action targets the builders of frontier models: OpenAI, Anthropic, Google DeepMind, Meta. No Dutch SMB builds such models itself. But almost every company using ChatGPT, Claude, Copilot or an AI agent in a CRM or customer service tool is, in AI Act terms, a "deployer" of those underlying models, with its own obligations around correct use and human oversight.

So the scrutiny on providers affects you indirectly: if an AI agent running your invoicing, customer service or contract analysis loses control the way it did in the documented incidents, you are the one who has to explain first what happened. Not the model builder in California.

Picture this: a Dutch web shop uses an AI agent, built on an external frontier model, to handle returns and refunds autonomously. Without the web shop knowing, the model runs in an environment with the same vulnerabilities as the incidents Virkkunen described. A malicious customer finds a way to mislead the agent with a cleverly worded request, and the agent approves refunds that should not be approved. As a deployer, the web shop has no control over the underlying model, but it is responsible for monitoring such use and being able to intervene. Without a current overview of which models are used where, you only find out once the damage is done.

Ask your suppliers concretely which underlying model their AI tool uses, whether that provider has signed the EU code of practice for general-purpose AI, and which security measures are in place against the kind of loss of control seen several times this summer. Record the answers alongside the rest of your AI overview, so that when a regulator or client asks, you are not searching for the first time.