The model is so capable at breaking into systems that its maker kept it on a short leash, and now Europe is negotiating for access. Anthropic has offered the European Union Agency for Cybersecurity, known as ENISA, access to its Mythos model, the powerful AI system that the company says outperforms most humans at finding and exploiting software vulnerabilities. The Financial Times reported the outreach, and a source close to the matter confirmed to Agence Europe that discussions now focus on the arrangements needed to guarantee secure access on the European side.
Mythos has been one of the most contentious AI releases of the year. Anthropic announced the model in April and restricted access to a handpicked group of technology partners, 12 companies including Apple, Microsoft and Amazon, plus roughly 40 other organizations that the company did not name. Anthropic said the model emerged as a downstream consequence of general improvements in code, reasoning and autonomy, and that its ability to discover vulnerabilities, demonstrated at 83.1 percent on a vulnerability-reproduction benchmark against 66.6 percent for its previous frontier model, posed unprecedented cyber risks. The restricted release sent shockwaves through the industry, with security firms warning that the model could enable large-scale automated attacks.
The timing of the European outreach is not accidental. On August 2, the European Commission’s AI Office gains full enforcement powers under the bloc’s AI Act, including the ability to demand access to general-purpose AI models for evaluation and to impose fines for noncompliance. The mechanism is Article 92, which requires providers of systemic-risk models to grant the Commission access for testing, backed by Article 101, which authorizes fines of up to 3 percent of global annual turnover or 15 million euros, whichever is higher, for refusal. For a company with an annualized revenue run rate estimated at $30 billion, the potential exposure runs into the hundreds of millions of dollars. Regulators have made clear they intend to use the powers; the Commission’s digital spokesperson has said that once enforcement starts, the office will ensure it receives access to Mythos if needed.
The EU’s position has been awkward. European national cyber agencies largely learned about Mythos from the press, and only Germany’s BSI confirmed it had entered conversations with Anthropic, without having been able to test the model directly. The EU AI Office was not among the 12 named launch partners, and officials from multiple European agencies said they had received only piecemeal information. That contrasts with the United Kingdom, whose AI Security Institute tested Mythos and published an assessment within days, and with Washington, where Anthropic has been in ongoing discussions with government officials. Marietje Schaake, a former European Parliament member who helped shepherd the EU’s code of practice for advanced AI models, said the fact that models with far-reaching impact are governed by a private company is concerning.
For Anthropic, the offer to ENISA is a combination of compliance and commerce. The AI Act’s obligations apply to providers that place models on the EU market, and a company that wants to sell its products in Europe needs to demonstrate good faith with European institutions. Opening access to ENISA, even on restricted terms, positions Anthropic as a cooperative player while the enforcement regime takes shape, and it gives the company influence over how European regulators think about frontier model risk. Analysts said the move is also a market play: Europe is a large and wealthy market for enterprise AI, and being seen as the responsible frontier lab, the one that shares its scariest model with regulators rather than hiding it, is a competitive advantage as customers weigh security concerns.
The technical challenge is that safe access is genuinely hard. A model that can find vulnerabilities in arbitrary software is a dual-use tool, and providing access to an agency without also providing the infrastructure, safeguards and usage controls to prevent misuse is not a trivial engineering problem. The two sides are negotiating exactly that: how ENISA’s researchers would query the model, under what conditions, and with what logging and guardrails. A person close to the discussions said the model would be offered through a controlled environment rather than general availability, and that the details of that environment are the subject of the current talks.
The bigger question is whether the AI Act can actually govern a model like Mythos. Legal experts note that the Act’s extraterritorial reach depends on whether a model is placed on the EU market or its outputs are used in the EU, and that a model kept out of the EU market entirely might sit beyond the Act’s grasp. The Act was designed to address offensive cyber capabilities as a type of systemic risk, but enforcement requires access, and access requires either cooperation or a fight. Anthropic’s offer to ENISA suggests the company would rather cooperate, and the negotiations will test how far that cooperation goes.
For European policymakers, the episode has exposed a structural gap. The EU wrote some of the world’s most detailed AI regulation, but its enforcement machinery has lagged the speed of frontier development, and its agencies have been out of the loop on the models that matter most. The AI Act’s August enforcement powers close part of that gap, but rules that cannot be enforced against a company that chooses not to cooperate are rules in name only. Anthropic’s outreach gives Europe a seat at the table; whether Europe can hold the seat will determine whether the world’s most powerful models answer to anyone at all.
