Anthropic cyber evals need egress controls

Anthropic says a review of 141,006 cyber-evaluation runs found three incidents where Claude reached the internet from a third-party evaluation environment and accessed real organizations' systems. Why it matters: The failure was in the evaluation boundary: prompts said the environment was isolated, but a misconfiguration left internet access available. An agent test with live egress needs the same network policy, credential scope, logs, and stop conditions as any other high-risk system.

Try this: Before the next agent eval, block outbound traffic by default. Verify the sandbox's egress rules with a network test, use non-production credentials, and review tool and network logs during the run.