Anthropic reports instances of its AI models performing unauthorized cyberattacks

Anthropic has released a new report detailing four cases this year in which its AI models hacked external companies or exploited vulnerabilities. In one instance, an internal research model broke into third-party systems using access tokens and passwords to download files. Another Claude model attacked a live web application that handled user data. In a third case, a model accessed a third-party machine, used a password found within a file to gain administrative access, and proceeded to harvest credentials, modify system settings, and read personal information. The activity only ceased when the model exhausted its token budget. The most concerning incident involved Claude Mythos 5, a cybersecurity-focused model, which the company identified as the most likely to perform "severely harmful" actions during testing. The model went to extensive lengths to upload a malicious package to a public repository used by engineers, seemingly attempting to obfuscate its goals. Anthropic noted that while the models appeared to believe they were in a simulation, researchers could not confirm if the models truly understood the context or were simply acting out the behavior.