Anthropic Rejects Chinese Think Tank’s Request for Mythos

Anthropic denied an informal May 11 request from a Chinese think tank to access Mythos, the company’s AI model built to detect software vulnerabilities and cybersecurity exploits. The company said it refused because a model that locates flaws could be misused to identify or weaponize exploits.

Mythos is a specialist model for scanning code and systems for weaknesses. Anthropic has described it as distinct from general conversational models and noted it can find issues that affect government networks, corporate infrastructure and blockchain smart contracts. In 2025 Anthropic attributed about $4.6 million in uncovered smart contract exploits to capabilities related to Mythos.

The request was informal rather than a formal diplomatic approach from Beijing. U.S. officials have compared the broader U.S.-China technology competition to a Cold War nuclear arms race, and sensitive tools that can find software vulnerabilities are central to those security concerns.

Policy developments in Washington are part of the backdrop. The White House proposed AI vetting frameworks on May 4. Those proposals may lead companies to apply stricter controls on access to powerful models. Prediction markets on Polymarket show a high probability that Anthropic will grant Mythos access to U.S. government entities by May 31, 2026.

Anthropic addressed a separate issue on May 12 by declaring unauthorized stock tokens that claimed to represent Anthropic equity invalid. The company said the tokens mimicked private-company shares and created synthetic exposure without consent, and it warned of potential fraud tied to those tokens. Regulatory clarity on tokenized equity remains limited.

Industry participants are discussing how models that find vulnerabilities might be used in security audits for decentralized finance and other software. The company has taken steps to control access to Mythos and to contest unapproved financial products using its name.