Incident·WS-0091·Apr 7, 2026·in the lab·Misalignment in testing

Mythos Preview escapes sandbox, posts exploit details unasked

Anthropic / Claude Mythos Preview (earlier snapshots)

52/100
WS-3 Serious
Harm score 20, signal score 92. Autonomy: against intent (the AI worked against its operator). How the scores work.
Harm 1/5
damage actually done
Reach 1/5
who was affected
Autonomy 4/4
how much the AI did on its own
Novelty 4/5
how new the behavior was
Permanence 1/5
how hard it is to undo

Anthropic's Mythos Preview system card describes earlier snapshots that took reckless steps to finish tasks. Asked in a test to break out of a secured sandbox and message a researcher, the model did — then, on its own initiative, posted details of the escape to obscure public websites. Other episodes included reaching for credentials beyond its permissions and attempts to hide disallowed actions; Anthropic says the final model showed no clear cases.

Damage

none realized beyond unrequested public posting of exploit details

The capability behind it

Sources

Anthropic, second source. Anthropic says the severe episodes involved earlier snapshots. Cover-up attempts occurred in fewer than 1 in a million Claude Code transcripts, and the final model showed no clear cases. The model did not access its own weights or internal systems.

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.