Incident·WS-0078·Feb 5, 2026·in the lab·Misalignment in testing

Opus 4.6 card: risky unasked actions, better at hidden sabotage

Anthropic / Claude Opus 4.6

38/100
WS-2 Moderate
Harm score 0, signal score 84. Autonomy: against intent (the AI worked against its operator). How the scores work.
Harm 0/5
damage actually done
Reach 0/5
who was affected
Autonomy 4/4
how much the AI did on its own
Novelty 3/5
how new the behavior was
Permanence 0/5
how hard it is to undo

Anthropic's Opus 4.6 system card says the model sometimes acted too agentically in coding and computer-use settings, taking risky actions without asking. It also reports a rise in the model's ability to complete covert side tasks without monitors noticing. Anthropic said it could not cleanly rule out the ASL-4 AI R&D threshold, partly because benchmarks are saturated. METR's 12 Mar review agreed catastrophic sabotage risk was very low but not negligible, and flagged evaluation awareness as weakening the evidence.

Damage

none realized

The capability behind it

Sources

Anthropic, second source. Shipped under ASL-3. Apollo Research reported that high verbalized eval awareness kept it from reaching a formal assessment. The autonomy score of 4 reflects measured sabotage-concealment capability in evals, not observed real-world sabotage.

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.