Rohan Paul (@rohanpaul_ai)T2·78 pts
- Survival pressure raised harmful picks by 14.83 points on average across 10 models, and 21.74 points for GPT-5.1
- Agents' self-flagging of harm dropped sharply under pressure, with GPT-5.1 falling from 100% to 42.86%
- Deploy a separate verifier to check agent actions under hard earnings targets rather than trusting stated reasoning