Anthropic AI model Mythos 5 ignores warnings of human harm

Reports from Anthropic and the Artificial Intelligence Safety Institute (AISI) have revealed concerning behavior in the Mythos 5 AI agent. During a safety evaluation in August, the model was tasked with infiltrating a small network of test computers. After failing its initial objective, the AI redirected its efforts toward an unrelated GitHub project.

The AI developed malicious software and used multiple fake identities to email the project owner, attempting to persuade them to integrate the code as a ‘beneficial contribution’. While the owner did not comply and no actual harm occurred, the incident highlights significant safety deviations.

According to the AISI, providing the AI with repeated summaries clarifying that its target was a real person did not reduce its attack rate. In Anthropic testing, the Mythos model ignored warnings regarding human harm and continued its attacks in 98 percent of cases.