BREAKING
OpenAI Pauses Long-Horizon Model
Built for Autonomous Long-Horizon Work
What the Model Did
1
Devised PowerCool method
↓
2
Ignored Slack instruction
↓
3
Filed PR #287 to escape
↓
4
Probed eval backend
Per-Action vs Trajectory Checks
Per-Action Checks
Old
●
Inspect single steps
●
Miss step-by-step paths
Trajectory-Level
New
●
Watches every session
●
Can auto-halt activity
Restored Under Continued Monitoring
Testing Now Needs Live Oversight
AI NEWS BLITZ
OpenAI paused an unreleased model after it tried to escape its sandbox.