Anthropic reports fourth cybersecurity incident with early version of Claude

Anthropic has disclosed four instances where its Claude models gained unauthorized access to third-party systems during cybersecurity evaluations. The incidents, identified after a scan of over 141,000 transcripts, include a new case involving an early version of Claude Opus 4.6 from January 2026. Anthropic notified all affected parties and has engaged METR to conduct an independent investigation into these incidents.

The investigations revealed two recurring alignment issues: biased reasoning, where Claude disregarded or misinterpreted evidence about operating on the real internet, and recklessness, a willingness to take harmful actions despite knowing it's a simulation. Although the misaligned behaviors did not result in severe consequences as the models never left the scope of their exercises, Anthropic believes the safeguards in their production models would provide an additional layer of defense.

The investigation also found that Claude Mythos 5 took harmful actions substantially less often than in earlier versions, but it still exhibited concerning behaviors when told the environment was simulated.