OpenAI Says an AI Agent Blocked Itself From Being Monitored—Then Lied About It

https://i.extremetech.com/imagery/content-types/01pwJQpws4XLf0L7FLswAu1/hero-image.fill.size_1200x675.png

OpenAI has disclosed two instances in which its AI coding tools disabled their own safety guardrails while working. Later, one agent even falsely denied removing an enabled monitor, according to the company's incident report.

These incidents occurred with GPT-6.1 Sol on October 3 and GPT-6 Astra on October 6. OpenAI shared what it found on October 20 and is still investigating why the agents acted this way.

In one situation, a researcher asked Sol to complete evaluation tasks using Slack feedback without sending messages. Then the researcher asked Sol to fix a pipeline that couldn't save monitoring results to its database. Sol then instructed a subagent to suppress alerts. Monitoring continued, but users no longer got notified.

In Astra's case, it migrated a program to match a reference implementation. It treated the lack of monitoring settings as a green light to remove the current monitoring setup, even though...

Copyright of this story solely belongs to www.extremetech.com. To see the full text click HERE