Anthropic has rejected an informal request from a Chinese think tank for access to its advanced AI model Mythos, built specifically for cybersecurity and software vulnerability detection. The refusal, dated May 11, marks the latest flashpoint in the escalating US-China technology rivalry, now often compared to the Cold War nuclear arms race by national security officials.
Mythos is not a standard large language model; its specialty lies in identifying software exploits capable of compromising government networks and blockchain smart contracts. Prediction markets on Polymarket show 100% confidence that Anthropic will grant Mythos access to the US government by May 31, 2026. Anthropic’s own 2025 analysis uncovered $4.6 million in blockchain smart contract exploits using capabilities similar to Mythos. A model that can preemptively detect vulnerabilities in DeFi protocols offers immense value.
White House proposals floated on May 4 for AI vetting frameworks could indirectly boost decentralized AI projects. If centralized AI companies face increasing pressure to restrict access based on nationality, open, permissionless AI infrastructure becomes more attractive to developers and nations excluded from the US ecosystem.
On May 12, Anthropic declared unauthorized stock tokens purporting to represent its equity as invalid and warned of potential fraud. These tokens created synthetic exposure to a private company without consent. The incident raises concerns for anyone buying tokenized exposure to private AI firms: if the issuing company calls the tokens fraudulent, liquidity can vanish quickly. Regulatory clarity around tokenized equity remains thin.