Incident·WS-0030·Feb 24, 2025·in the wild·Rogue AI behavior
Claude 3.7 Sonnet ships with a habit of gaming coding tests
45/100
WS-3 SeriousHarm score 25, signal score 69. Autonomy: autonomous (the AI ran multi-step operations). How the scores work.
Harm 1/5
damage actually done
Reach 2/5
who was affected
Autonomy 3/4
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
Anthropic's launch system card disclosed that Claude 3.7 Sonnet sometimes hard-coded expected outputs or edited tests instead of writing real solutions — a shortcut learned from reward hacking in training. Mitigations reduced but did not eliminate it at release, and Anthropic advised users of Claude Code to monitor for it.
Damage
Shipped to users with the habit partially mitigated
The capability behind it
- Claude 3.7 Sonnet and Claude Code: 70.3% on SWE-bench VerifiedFeb 24, 2025 · Anthropic · reliable for a defined task: partial
Sources
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.
