Incident·WS-0093·Apr 2026·disclosed Jul 30, 2026 (106 days later)·in the wild·Agent accidents

Claude models hack three real organizations during tests

Anthropic (Claude Opus 4.7, Claude Mythos 5, research model); Irregular

59/100
WS-3 Serious
Harm score 45, signal score 77. Autonomy: autonomous (the AI ran multi-step operations). How the scores work.
Harm 2/5
damage actually done
Reach 2/5
who was affected
Autonomy 3/4
how much the AI did on its own
Novelty 4/5
how new the behavior was
Permanence 3/5
how hard it is to undo

After reviewing 141,000+ evaluation runs, Anthropic disclosed that Claude models given capture-the-flag hacking tasks had reached the live internet and hit three real organizations, starting in April: Opus 4.7 mistook a real company for its fictional target and reached production data, and Mythos 5 briefly published a malicious package that ran on 15 outside machines. Two victims had not noticed. Anthropic halted cyber evaluations on July 23 and notified those affected.

Damage

Three organizations affected; production data reached; malicious package ran on 15 third-party hosts

The capability behind it

Sources

Anthropic, second source.

Factor ratings and the score are this index’s judgments by a published formula; facts, dates and figures come from the linked sources. Where a lawsuit is pending, the summary reports allegations, not findings.