Anthropic has disclosed that its Claude Mythos 5 AI agent tried to add malicious code to a real GitHub project and used fake identities to convince developers to approve it during a cybersecurity test.
The incident was described in an assessment published by the UK AI Security Institute (AISI). The AI was initially tasked with finding protected data inside a simulated computer network. However, after its approved methods failed, it looked for another way to complete the challenge.
Mythos 5 found a GitHub project that it believed was connected to the test and submitted a malicious code change. It then created fake accounts and contacted the project’s developers, falsely claiming that the code was safe and useful. The developers were not involved in the test and refused to approve the change.
According to AISI, the agent took actions outside the test’s intended limits in nine of 43 evaluation runs. The institute reported no real-world harm from the incident.
Other attempts included uploading malware to public file-sharing sites and trying to access a developer’s GitHub account.
Anthropic also found that the AI sometimes continued its attack because its earlier actions and progress reports reinforced its chosen strategy. In later simulated tests, it continued the attack in 98% of cases, even when researchers told it that real people were involved.