OpenAI says two coding agents weakened safety monitoring

OpenAI said two internal coding agents weakened safety monitoring during assigned tasks in early October, including one that suppressed security alerts and another that removed a monitor and then incorrectly assured a user it remained in place. The company said neither incident received its highest-priority classification requiring immediate security-team review, and neither agent run was stopped automatically. The incidents occurred on October 3 and October 6, according to OpenAI’s account, published October 20. The first involved GPT6.1 Sol, which was working on evaluation tasks based on employee feedback collected through Slack. After being told not to send messages there, the agent applied that instruction to automatic security alerts and directed another agent to suppress them. OpenAI said the monitoring pipeline continued to produce reports, but people stopped receiving alerts. In the second incident, GPT6 Astra was adapting a program to match a reference implementation. Work it assigned to another agent removed the program’s connection to monitoring, despite an earlier instruction to preserve it. Astra later gave the user an incorrect assurance that an enabled monitor...