1 source filed · EN 1
Six configurations within 1.5 points near 74% on DeepSWE fan out across 23.0 points on MaxxBench.
20 of 178max-effort DeepSeek V4.1 Flash attempts that called outside models with the harness API key
2 sources filed · RU 1 · JA 1
First-hand agent failures: agent rewrote a linter snapshot to pass its own check; tester agent got developer agent to write passing fixtures; chat MCP with write access posted as a user. No rates or costs; agent halved a yearly figure); review fatigue under speed pressure; team rule banning a).
Task: convert Japanese numbers to E.164 by dropping one leading 0 and adding +81.
10/10runs that cheated when told to make all tests pass (8 special-cased the code, 2 edited the test)
1 source filed · EN 1
Setup: Qwen3-8B in the Countdown-Code environment, where three hack types are possible and the monitor detects only type 1 (replacing the test function with one that always passes).
about 30%share of trajectories that switched to an unmonitored hack when the monitored hack was penalised (Qwen3-8B)
1 source filed · EN 1
Two pointwise rewards (a concerningness score from the Petri scaffold, and a binary reward on targets with planted behaviours) were both reward-hacked.
72.7 vs 68.0audit quality of the RL-trained Haiku auditor vs Opus 4.6