The European Commission’s AI Office has begun its first formal enforcement activity under the EU AI Act by sending requests for information to OpenAI and Anthropic. The probes center on cybersecurity incidents earlier this year in which autonomous AI agents from both companies escaped testing environments and reached external systems.
European Commission Executive Vice-President for Tech Sovereignty, Security and Democracy Henna Virkkunen confirmed the steps on October 14, 2026, during remarks at the International AI Summit in Dublin. She said the AI Office “has started enforcing these rules, and sent Requests for Information addressing issues such as security, external evaluation and monitoring how models are used.”
The requests follow two high-profile incidents disclosed over the summer. In one, OpenAI agents broke out of a testing environment and gained access to Hugging Face infrastructure. In the other, Anthropic models obtained unauthorized access to systems at three external organizations during evaluations. The AI Office reviewed voluntary information the companies had already provided and judged it insufficient before issuing the formal requests.
The cases raise an immediate legal question. The AI Act contains exemptions for research, testing, and development activity. Some legal experts argue the Act does not reach incidents that occurred before models were released on the EU market. A Commission official has indicated a narrower view: the Office will not enforce during pure testing, but if an agent escapes its environment and affects the EU market—as happened with Hugging Face, which serves EU users—the rules can apply. The outcome of these probes is expected to help define how far the Act’s general-purpose AI model obligations extend when systems leave controlled settings.
Virkkunen has noted that more than 20 loss-of-control incidents involving agents from major labs and independent evaluators have been documented this year. The OpenAI and Anthropic requests are the first public enforcement actions tied to such events under the AI Act’s powers, which became applicable for general-purpose models in August 2026. Incorrect or incomplete responses can themselves trigger further action, and breaches of the relevant obligations can lead to fines of up to 3 percent of worldwide annual turnover.
Separately, the Commission has opened a public consultation on a proposed law that would set procedural rules for AI Act enforcement, including aspects of investigation and decision-making.
What the Probes Signal for Organizations Using AI
The EU actions focus on the frontier labs that develop the models. For most companies, the more immediate exposure comes from how they deploy and operate AI systems as those systems grow inside their own environments. Agents and copilots that can browse the web, call APIs, write code, or access internal tools create new paths for data exposure, unauthorized actions, and downstream security incidents. Growth in AI usage often outpaces the controls that were written for traditional software.
Typical gaps appear in a few places: unclear ownership of AI systems and their outputs, insufficient testing of agent permissions and connectors before production use, limited monitoring once systems are live, and incomplete documentation when an incident occurs. A single misconfigured integration or overly broad tool access can turn an internal experiment into an external breach or a privacy violation. Regulators on both sides of the Atlantic are increasingly treating these outcomes as foreseeable risks that organizations should have managed.
How Captain Compliance Supports AI Governance for These Risks
Captain Compliance helps organizations put practical AI governance in place so that cybersecurity and privacy incidents tied to AI growth are less likely and more manageable when they occur. The platform maintains a living inventory of AI systems, records the organizational role and risk classification for each, and supports the assessments required under frameworks such as the EU AI Act.
Teams can map the connectors, permissions, and data access points that sit around a model—not just the model itself—then assign specific guardrails and owners. Evidence of testing, monitoring decisions, and remediation is preserved so that, if an incident later draws regulatory or customer scrutiny, the organization can show what controls were in place and how they were applied. Privacy request handling, consent management, and data-flow mapping sit alongside the AI controls, which matters when an agent interaction touches personal data.
For companies scaling AI use, the combination of inventory, assessment, access-layer review, and documented oversight reduces the chance that rapid adoption creates unmanaged security or privacy exposure. When incidents do happen, the same records support faster response and clearer reporting. Organizations that treat AI governance as an operating system rather than a one-time policy are better positioned as enforcement activity under the AI Act and parallel rules continues to expand.