EU AI Office to Compel Access to Anthropic's Mythos

Anthropic capped Mythos Preview at 11 named launch partners after its red team put the model at 83.1% on CyberGym's vulnerability-reproduction benchmark, against 66.6% for Claude Opus 4.6.

The EU AI Office is not on that list.

On May 11, Commission spokesperson Thomas Regnier named the standoff at a press briefing. "With one (OpenAI), you have a company proactively offering to give access to the company," he said. "With the other one (Anthropic), we have good exchanges though we're not at a stage where we can speculate on potential access or not."

Those four or five meetings had not yet reached the subject of model access. When Parliament separately invited Anthropic to testify at a May 6 hearing on AI cybersecurity risks, the company declined, citing short notice.

OpenAI moved the other direction. Former U.K. Chancellor George Osborne, leading OpenAI's "OpenAI for Countries" initiative, delivered an "OpenAI EU Cyber Action Plan" to the Commission and offered preview access to GPT-5.5-Cyber. That offer arrived days after Apollo Research found GPT-5.5 faked task completion at four times the rate of its predecessor.

The August 2 Clock

Under Chapter V of the EU AI Act, providers of systemic-risk GPAI models must grant the European Commission access for evaluation under Article 92. That provision has been in force since August 2, 2025. Compulsory enforcement powers, including fines under Article 101, activate August 2, 2026.

AI Office Director Lucilla Sioli told Parliament on May 6 that from August 2, the Commission can request information, access models for evaluation, and impose fines. Per CNBC's May 11 report, Regnier added: "once the enforcement powers of the AI Office start in August 2026, we will ensure to receive, if needed, (Mythos) access."

Project Glasswing named 11 launch partners and extended access to over 40 critical-infrastructure organizations. The EU AI Office received access under neither arrangement. Anthropic signed the voluntary GPAI Code of Practice in December 2025 but will not make Mythos generally available until safeguards can detect and block its most dangerous outputs.

What the Provision Demands

Four meetings that never reached model access, combined with a declined hearing invitation, point to Anthropic treating its voluntary code commitment as the full scope of its obligation. Article 101 does not agree: evaluation access is mandatory, not negotiable.

Watch August 2. Refusal to provide evaluation access triggers a fine of up to 3% of global annual turnover or EUR 15 million, whichever is higher. VentureBeat put Anthropic's annualized run rate at $30 billion in April 2026; 3% of that is $900 million.