Vulnpocalypse: AI that finds software vulnerabilities could magnify cyber risks

Anthropic said it will withhold its Mythos Preview model because of its unusually strong ability to find software vulnerabilities; U.S. officials and industry leaders have begun discussions, including a meeting convened by the Treasury Secretary.

Anthropic announced it will not publicly release its latest model, Mythos Preview, citing unusually strong capabilities at discovering software vulnerabilities. The company said it is sharing the model only with a limited set of tech partners to help address those risks. U.S. officials and industry leaders have taken note, and a meeting was convened by Treasury Secretary Scott Bessent with major financial institutions. Security researchers warn similar capabilities could spread in the near future and affect how cyber threats are discovered and exploited. Key developments: - Anthropic withheld Mythos Preview from public release, citing its vulnerability-discovery capabilities. - The company is sharing the model with a limited group of technology partners and vendors. - Treasury Secretary Scott Bessent convened a meeting with major financial institutions to discuss rapid AI developments, an agency spokesperson said. - Industry experts reported that comparable models could become more broadly available within six to twelve months. Summary: The announcement has prompted government and industry discussions about how to manage AI systems that can rapidly find software flaws. Experts say this may increase the risk that less-skilled actors can exploit vulnerabilities. Further coordinated responses and outcomes are undetermined at this time.