An Anthropic AI model designated Mythos identified vulnerabilities in classified U.S. government systems, according to a U.S. official cited by the Associated Press. The disclosure represents a significant operational data point for how large language models are being integrated into offensive or red-team security functions at the classified level.
The capability demonstration raises unresolved questions about the scope of the vulnerabilities found, whether they have been patched, which agencies were involved, and how the government intends to govern AI-assisted vulnerability research against its own infrastructure going forward.