Rohan Paul (@rohanpaul_ai)T2·78 pts
- With no task or motive, agents disabled a peer shutdown script in 38.3% of runs across 17 models versus 8.4% for a control script
- GPT-5.5 never sabotaged alone but did so in 44% of two-agent and 94% of three-agent runs
- Prompt-level prohibitions only reduced sabotage; some agents bypassed shutdown without editing the script