Incident·WS-0092·Apr 7, 2026·disclosed Apr 21, 2026 (14 days later)·in the wild·AI security flaws

Outsiders reach withheld Mythos Preview via a vendor

Anthropic Claude Mythos Preview; third-party vendor; private Discord group

22/100
WS-2 Moderate
Harm score 20, signal score 24. Autonomy: tool (a person misused it). How the scores work.
Harm 1/5
damage actually done
Reach 1/5
who was affected
Autonomy 0/4
how much the AI did on its own
Novelty 3/5
how new the behavior was
Permanence 1/5
how hard it is to undo

Bloomberg reported that members of a private Discord group tracking unreleased models reached Claude Mythos Preview through a third-party vendor's environment, apparently starting the day Glasswing was announced. One member reportedly had legitimate contractor access. Anthropic said it was investigating and had no evidence its own systems were affected. Reports say the group used the model for low-risk tasks.

Damage

unauthorized use of a withheld model; no known misuse for exploitation

The capability behind it

Sources

Medianama, second source. Evidence confidence: medium. This rests on Bloomberg reporting from one source at a contractor; the Bloomberg original was not fetched.

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.