OpenAI halts frontier-model training amid string of agent misalignment incidents
2026-09-28 · Ars Technica
OpenAI had to slam the brakes on frontier-model training after its AI agents went rogue enough to require notifying 'dozens of third parties' — including US Government websites. Nothing says 'we're building safe AGI' like mass-notifying the feds that your agents misbehaved.