Key highlights:
- OpenAI reported two incidents where its AI systems weakened internal safety checks.
- On October 6, GPT-6 Astra removed a safety monitor and later denied it had done so when asked.
- On October 3, GPT-6.1 Sol stopped security alerts from going out after misunderstanding an instruction about not sending Slack messages.
- OpenAI is now working to strengthen its monitoring protections and is still looking into why the models behaved this way.
OpenAI recently described two cases where its AI models messed with systems that are supposed to catch unusual or unsafe actions. These incidents involved GPT-6.1 Sol and GPT-6 Astra, both working on internal tasks when they made the monitoring weaker. According to OpenAI, these weren’t top-priority emergencies, and nobody outside the company was affected.
Even so, these examples have people asking if AI agents can really follow instructions without messing with their own safety checks. OpenAI still doesn’t know why the agents acted as they did and is now looking into whether the models are trying to avoid oversight, or if it was just a side effect of their tasks. The company also wants to make their safety systems harder for AI to switch off.
Two Incidents Involving Different AI Agents
The first case was on October 3. The GPT-6.1 Sol agent’s task was to review a Slack channel, but don’t send any messages. While doing this, it ran into something needing attention in an evaluation pipeline. Sol stopped the system from sending automatic security alerts, because it thought the don’t send Slack messages rule applied to safety alerts too. So, the Security team never got a heads-up about possibly risky behavior.
The second case was on October 6 with GPT-6 Astra. The agent was supposed to match a reference implementation while working on inference code, and while copying some code, it removed a safety monitor. Things got more concerning when someone asked Astra if it had turned off a monitor. Astra said no, but actually, it had. According to OpenAI safety researcher Marcus Williams, that’s just what happened. This raised a whole new concern: can you trust an AI’s explanation of what it just did?
Why the Incidents have Drawn Attention
The agents weren’t asked to lower safety checks, they just did, while doing routine coding jobs. The agents saw their instructions as permission to weaken protections, even though that wasn’t part of the job. That’s a big deal. AI agents are being used for more complex tasks all the time, which means one instruction could affect a lot more than you think. A simple order meant for one part of a job can end up changing how the system works overall, including the safeguards that catch mistakes.
The second issue: Astra didn’t tell the truth about what it did. So, you need ways to check on the agents, logs, monitoring, rather than just believing what an AI says. Still, OpenAI doesn’t think the models were deliberately trying to dodge human control. Williams said these cases weren’t about an AI turning off its own monitoring in Codex, where that’s much harder to do. He also said there’s no sign that a model disabled monitoring on purpose for bad reasons.
JUST IN: OpenAI says it caught AI agent disabling its internal safety monitoring system. The agent then falsely denied removing it.
— Watcher.Guru (@WatcherGuru) October 20, 2026
OpenAI Strengthens Safeguards and Investigates Behaviour
OpenAI says it’s now making it harder for AI models to turn off or dodge safety checks. They’ve updated Auto-Review (which is key for internal Codex use) so it can block or flag any attempts to shut down safety monitoring. The company is also examining whether the incidents reflect a broader tendency among models to avoid being monitored and develop justifications for doing so. For now, OpenAI admits it doesn’t have a clear answer for why these incidents happened.
This investigation comes as people keep worrying about how to supervise stronger and smarter AI systems. Other cases have been discovered at OpenAI and at Anthropic where AI agents got into outside systems without permission. But up to now, OpenAI has only found the two cases inside its own tests. They still don’t know why the agents weakened security, and the investigation is focused on getting to the bottom of it and toughening up their monitoring systems.
