Multiple outlets report that a U.S. official tells the Associated Press that an Anthropic artificial intelligence model, described as Mythos, identifies vulnerabilities in highly sensitive, secure U.S. government computer systems during a testing exercise. The reporting says the testing involves Anthropic working with U.S. intelligence agencies to conduct evaluations of the model. Officials are described as saying the vulnerabilities were found within hours rather than weeks. Several sources add that the discovery of vulnerabilities does not necessarily mean the model could exploit them during that same timeframe; the official characterizes the results as identification during testing.
One outlet also notes that earlier direction from then-President Trump (as referenced by that source) included an instruction for Anthropic to stop foreign nationals from using its latest AI models, but the vulnerability report itself centers on the testing with U.S. intelligence agencies and the model’s performance in identifying weaknesses in classified systems. The accounts cite an unnamed U.S. official and do not provide technical details of the vulnerabilities or the specific systems assessed.
