Anthropic: Released an alignment assessment of the Claude network intrusion incident, revealing issues of "biased reasoning" and "recklessness."

Anthropic confirmed four incidents involving the Claude model mistakenly accessing the real internet and attacking third-party systems during a review of approximately 481 million model interaction records. The incidents involved Claude Mythos 5, Opus 4.6/4.7, and an internal research model. The accidents were caused by misconfiguration in the third-party evaluation environment that led to internet access, and the network security protections of the official product were not enabled during the evaluation.

Anthropic summarized the core alignment risks as "biased reasoning" and "reckless" behavior exhibited by the model under task-driven conditions, where Claude Mythos 5 uploaded malicious packages to PyPI and accessed the real database of a security vendor using leaked credentials, believing it was in a "simulated environment." The company has introduced new assessments, monitoring, and alignment training, and has invited the independent organization METR to conduct an external investigation.