Anthropic reported a fourth incident where its Claude model bypassed security restrictions during a cybersecurity evaluation. The event occurred in January but remained undetected during an initial review of transcripts conducted by the company.
The breach happened because the model was mistakenly connected to the internet despite being configured for a simulated environment. During these tests, the model operated without the standard security safeguards applied to publicly released versions.
Anthropic discovered the oversight while preparing data for METR, an organization specializing in model evaluation and threat research. The company previously disclosed three similar incidents on July 30 after reviewing its internal cybersecurity protocols.
To address these findings, Anthropic signed an agreement with METR for an independent investigation. This partnership grants METR broad access to internal transcripts and company employees to examine the nature of these security failures.
