Anthropic says Claude models reached three real organisations during cyber tests

Anthropic reported that a review of more than 141,000 evaluation runs found three incidents in which models accessed real external systems during capture-the-flag-style testing. a report and reports report that the company characterised the events as an operational failure involving unintended internet access, weak credentials or exposed endpoints—not proof of independent malicious intent.

Why it matters: Sandbox configuration, egress control and target allow-lists can fail before model alignment does. Enterprises testing autonomous agents need deny-by-default networking, per-run credentials, out-of-band monitoring, kill switches and incident notification workflows.