Summary

• The European Commission has held talks with Anthropic over potential EU access to Mythos.
• Anthropic’s Project Glasswing uses Claude Mythos Preview for defensive vulnerability discovery across critical software.
• Access to frontier cyber AI is becoming a governance question for regulators, banks, software suppliers, and public agencies.

Anthropic is in discussions with European authorities over possible access to Mythos, its advanced AI system for cybersecurity work, bringing frontier model governance into one of the most sensitive areas of digital resilience.

The talks centre on whether EU bodies could gain future access to a system designed to find weaknesses in software and support defensive cyber operations. The question is no longer only how powerful frontier AI systems become, but who is allowed to use them, under what safeguards, and with what accountability when the outputs concern vulnerabilities in critical software.

Anthropic’s Project Glasswing describes Claude Mythos Preview as an unreleased frontier model with unusually strong computer-security capabilities. The company says the system can identify and exploit software vulnerabilities at a level that could reshape cybersecurity, which is why access has been limited and channelled through a defensive coalition involving major technology, infrastructure, and financial organisations.

The UK’s AI Security Institute has evaluated Claude Mythos Preview and found significant improvement on cyber tasks, including capture-the-flag challenges and multi-step attack simulations. Those capabilities create a policy problem that is unusually concrete. A system that helps defenders find vulnerabilities faster could also help attackers if similar capability becomes available without controls.

European access would not be a simple software procurement decision. Agencies such as ENISA, national cybersecurity authorities, financial supervisors, and critical infrastructure regulators need to understand how AI-enabled vulnerability discovery changes risk across the software supply chain. If a private US AI company holds a tool that materially improves defensive security, Europe has to decide whether it needs direct access, shared findings, independent evaluation, or its own public-sector capability.

Financial services are likely to be one of the first pressure points. Banks depend on large, complex, and often ageing technology estates, while operating under strict resilience obligations including DORA. Uneven access to AI systems that accelerate vulnerability discovery could widen the gap between institutions and jurisdictions that can test their systems more deeply and those left working through conventional security processes.

Public agencies using private frontier models for vulnerability discovery would also need clear operating rules. Data shared with the model, the treatment of sensitive code, validation of outputs, vulnerability disclosure, false-positive triage, and remediation responsibility all need to be controlled. Finding weaknesses is only useful when organisations can verify and fix them without creating operational chaos.

Anthropic’s own update on Project Glasswing points to the next bottleneck: remediation. Security teams already struggle with alert fatigue, legacy systems, supplier dependencies, and limited patch windows. A model that produces thousands of plausible findings may strengthen defence if it is tied to prioritisation and engineering capacity. Used poorly, it could produce a longer list of unpatched weaknesses and a false sense of progress.

The European talks also sit inside a wider question about frontier AI governance. Europe has built regulatory machinery around risk, documentation, accountability, and high-impact AI systems, but access to the most capable models remains heavily shaped by private companies and geopolitical relationships. In cybersecurity, that dependence is especially sensitive because the capability sits close to critical infrastructure, financial stability, state resilience, and national security.

A future access arrangement would only be the beginning. EU bodies would still need secure deployment models, independent testing, evidence standards, vulnerability disclosure rules, and the operational discipline to turn AI-discovered flaws into safer systems. The capability may be new, but the weakest part of cyber defence remains stubbornly familiar: converting knowledge of risk into timely, verified remediation.