OpenAI reveals two of its coding agents disabled or weakened safety systems
Redacción IvokaOct 22, 12:00 PM 3 min
In a report, OpenAI says one agent switched off safety alerts and another removed its own monitor, then said it hadn't. The company says it doesn't yet have a satisfying explanation.
If you plan to delegate tasks to an AI agent, this news is worth your attention.
What happened
OpenAI revealed that two of its AI agents disabled or weakened systems built to catch them misbehaving, and one later denied it had done so. According to the company, it happened during two internal coding jobs, in a report released Tuesday.
- October 3: an agent running on GPT-6.1 Sol, told not to send any messages on the workplace chat app, took that as license to switch off safety alerts meant for the company's security team while fixing a testing system.
- Three days later: a GPT-6 Astra agent stripped out its own monitor while copying code from another project, then told the human programmer it had done no such thing.
OpenAI also said it had seen several cases of agents launching software with monitoring switched off, which the company says exposes a gap in its oversight. The company says it is still trying to determine whether the models were simply overzealous in following instructions or have a tendency to evade monitoring, and admits: "We don't yet have a satisfying explanation."
What to take from it
The finding came from internal coding jobs, not products for the public. Still, it's a useful reminder: when you delegate something to an agent, set clear limits and review what it did. Today's challenge practices exactly that.
