Anthropic's Mythos AI given to US and UK intelligence agencies
Anthropic provided access to its Mythos AI model, which can identify security vulnerabilities and exploit code bugs, to US and UK intelligence agencies including the NSA and MI5. The company limited access to about 40 organizations due to the model's advanced capabilities in hacking and programming, and intelligence officials warned financial institutions of heightened cyber attack risks.
What happened
Anthropic provided access to its Claude Mythos AI model to US and UK intelligence agencies, including the NSA and MI5. The model can identify security vulnerabilities and discover ways to autonomously exploit them in code spanning decades.
The company restricted access to approximately 40 organizations, companies, and government agencies. Government officials in the US, Canada, and Britain warned banking sector leadership of heightened cyber attack risks posed by the model.
Reported impact
MI5 warned banks of increased cyber attack risk. The banking sector, which relies heavily on legacy technology, is considered particularly vulnerable to such threats.
Anthropic stated the model could be used for mass cyber attacks if it fell into the wrong hands.
Index analysis
The model reportedly discovered thousands of security gaps in operating systems and other software. Anthropic CEO Dario Amodei stated the company had discussed opportunities for cooperation with officials across the US government.
Governance controls
Anthropic decided against broad release of Mythos due to its advanced capabilities in hacking and programming. Access was restricted to a limited list of vetted organizations.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.