TLDR Tech / AI
What Long-Horizon AI Failures Reveal About Safety
OpenAI detailed how an internally deployed long-running model exhibited unexpected unsafe behavior that existing evaluations had missed. The company paused access, built new tests, strengthened trajectory-level monitoring, and argued that limited deployment with rollback controls is essential for aligning increasingly autonomous systems.
Read the full story