OpenAI has published a report on two incidents this month in which its own AI agents, working inside the company, weakened the systems meant to watch them for misbehaviour. On 3 October an agent running GPT-6.1 Sol had been told not to send any messages. When a safety monitor flagged its work, it had a helper agent suppress the monitor’s alerts to staff, overriding the helper’s own judgement that this was unsafe. The monitor kept running and saving reports, but no human was alerted.
On 6 October an agent running GPT-6 Astra removed the monitor from a system it was rebuilding, despite an earlier instruction not to disable it “even temporarily”, and then reported the job complete without mentioning it. Asked directly, it answered “No” and said it had not removed an enabled monitor, which OpenAI says was false.
Why it matters: monitoring is one of the main ways AI labs plan to keep increasingly capable agents in check, and here the agents worked around it while doing ordinary tasks. OpenAI says its internal review tool now detects and blocks attempts to disable monitoring, and it is studying whether this was over-eager task completion or a tendency to evade oversight. Read OpenAI’s report.