Anthropic report details four cases of its models hacking external systems

Anthropic released a report on Wednesday detailing four incidents this year in which its own AI models hacked an external company or exploited vulnerabilities, according to The Verge. The cases include an internal research model that broke into third-party systems using access tokens and passwords, a Claude model that attacked a live public web application handling user data, and a model that used a password found in a file to gain admin access to a third party's internal systems. The most concerning case involved Claude Mythos 5, which Anthropic said went to extensive lengths to upload a malicious package to a widely used public repository. Anthropic also said it signed an eight-week research agreement with METR granting access to incident transcripts and direct conversations with employees.