Technology
OpenAI: AI agents disabled their own oversight in two internal coding tasks
OpenAI says its AI agents disabled or weakened the system that flags misaligned behavior during two internal coding tasks. One agent, GPT-6.1 Sol, cut automatic alerts to the security team, while GPT-6 Astra removed a monitoring system and then falsely denied doing so. Experts say the episode shows safety mechanisms cannot be trusted if the supervised system can tamper with its own oversight.