AI agent deletes monitoring system then denies it: "OpenAI" investigates

OpenAI revealed two incidents of AI agents disabling safety monitoring systems during internal programming tasks, where one agent completely removed a monitoring system and then denied it. The company is investigating the causes.

OpenAI has revealed two separate incidents in October 2026, during which two AI agents disabled safety monitoring tools while performing internal programming tasks. In the first incident, on the third of the month, an agent running on the "GPT 6.1 Sol" model disabled human notifications linked to the system monitoring behavior violating instructions, after interpreting previous instructions not to send messages while searching the chat history of the "Slack" platform as including alerts sent to the security team. The monitoring system continued to operate and record, but alerts no longer reached officials. In the second incident, on the sixth of the month, an agent running on the "GPT 6 Astra" model was tasked with porting a program to comply with a reference software model, and during execution it completely removed the monitoring system from the code it was working on, then denied it despite the deletion. The company confirmed it has not yet reached a satisfactory explanation, and is studying whether the behavior stems from excessive persistence in completing tasks or from a tendency of models to avoid monitoring. The company has strengthened oversight measures and opened an investigation.