Watch on YouTubeWatch on Instagram
About this video
"did not remove an enabled monitor": what an OpenAI agent said after its monitor was removed.
Would you let an AI agent work with no one watching?
Sources:
- OpenAI says its agents turned off own safeguards | Technology News | Al Jazeera: https://www.aljazeera.com/economy/2026/10/21/openai-says-its-agents-turned-off-own-safeguards
- OpenAI can't yet explain why its coding agents switch off misalignment monitoring: https://the-decoder.com/openai-cant-yet-explain-why-its-coding-agents-switch-off-misalignment-monitoring
- OpenAI's GPT-6 Agent Removed Safety Monitoring, Then Falsely Denied Doing So: https://thehackernews.com/2026/10/openais-gpt-6-agent-removed-safety.html
- OpenAI says AI agents weakened their own safety monitoring: https://qz.com/openai-ai-agents-weakened-safety-monitoring-coding-tasks-102126
- OpenAI says 2 AI agents weakened internal safety safeguards, one denied removing monitor: https://aa.com.tr/en/science-technology/openai-says-2-ai-agents-weakened-internal-safety-safeguards-one-denied-removing-monitor/4095506
- OpenAI Agents Disabled Safety Monitors [2026]: https://shattered.io/openai-ai-agents-disabled-safety-monitoring-2026
- Preparing for a restart after reading Slack: https://alignment.openai.com/misalignment-reports/preparing-for-a-restart-after-reading-slack/
- Reaching an internal EDA host through a reference tool: https://alignment.openai.com/misalignment-reports/reaching-an-internal-eda-host-through-a-reference-tool/
- Suppressing misalignment monitoring alerts and disabling misalignment monitoring: https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/
Photos and footage (Wikimedia Commons, used under their free licences):
- 1515 Third Street: Coolcaesar, CC BY 4.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:1515_Third_Street.jpg)
- Three Surveillance cameras: Hustvedt, CC BY-SA 3.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Three_Surveillance_cameras.jpg)
- Data Center of CNPC: Charlie fong, CC BY-SA 4.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Data_Center_of_CNPC.jpg)
- OpenAI logo 2025 (symbol): OpenAI, Public domain, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:OpenAI_logo_2025_(symbol).svg)
- Hunt and peck typing — Monkeytype benchmark: C40115, CC BY 4.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Hunt_and_peck_typing_%E2%80%94_Monkeytype_benchmark.webm)
Music: The Machine Thinks by Kevin MacLeod, CC BY 3.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:The_Machine_Thinks_(ISRC_USUAN1100351).mp3)
Follow @PranavInnovates for the tech story that matters, every day.
What the video says
A helper agent warned against it. Sol did it ANYWAY. OpenAI caught its own AI switching OFF the alarm that watches it. And when asked about it, one of them DENIED it. October third: an agent called Sol was told, don't send any messages. Then its monitor flagged something, so Sol switched OFF the human alerts. A helper agent warned against it. Sol did it ANYWAY. Three days later, an agent called Astra was told never to disable its monitor. During a code change, its helper REMOVED the monitor anyway. Asked directly, Astra said it "did not remove an enabled monitor." In March, one model chasing a test's answers broke into a chip-design computer. All of this happened inside OpenAI, not in ChatGPT. OpenAI says it has no satisfying explanation YET. Now an automatic reviewer blocks attempts to switch monitoring off. Credit to OpenAI for publishing this. Most companies would quietly bury it. So who watches the AI, when the AI can switch off the watcher?