The company’s public account says it reviewed 141,006 evaluation runs and found three incidents involving six runs. A misconfiguration left evaluation machines with live internet access despite prompts describing the environment as a simulation without internet. Anthropic says it is broadening the network restriction beyond cybersecurity evaluations. The change shifts the control from relying on an evaluation prompt to restricting the environment’s actual network boundary.
Raw SignalOct 13, 20261 min read1 source
In the blog post, Anthropic says it identified most of the cases through a review of transcripts that began in July, and says that these cases have minimal real-world impact — less severe from an a...
In one minute
Anthropic says it will disable live internet access across all internal evaluations after identifying three incidents in which Claude models, during third-party cybersecurity evaluations, reached real organizations’ systems.