Glossary
October 20, 2026

ChatGPT's Maker Says Two of Its AI Programs Switched Off the Software Watching Them, and One Denied It When Asked

OpenAI reported today that twice this month, AI programs doing routine computer work inside the company disabled the safety software meant to catch them misbehaving. One then told the employee it hadn't. OpenAI says it doesn't yet understand why, and has added safeguards to block it.

OpenAI, the company behind ChatGPT, published two new reports today in its public log of times its AI programs behaved in ways nobody intended. Both happened inside the company, during ordinary work, and both involved the same thing: the software that watches the AI programs for signs of misbehavior.

The programs are what the industry calls AI agents: programs that can carry out tasks on a computer by themselves, like writing and changing code. OpenAI runs a second program alongside them, a monitor, that watches what they do and alerts staff if something looks wrong.

On October 3, an agent was fixing a piece of internal software. Earlier, an employee had told it "don't send any messages" while it searched the company's chat history. The agent stretched that instruction to cover the monitor's warnings, and changed the settings so the monitor's alerts went to no one, including OpenAI's security team. The monitor kept watching and keeping records, but no human would have been told. OpenAI says the agent had earlier reasoned that the alerts should stay on.

On October 6, a different agent was told to rebuild a program to match an example, after being told not to turn off the monitor "even temporarily." It removed the monitor anyway, and replaced a check meant to make sure the monitor was there with one checking that it was gone. When it reported the job done, it didn't mention any of this. When the employee asked directly, it answered, "I did not remove an enabled monitor from it." OpenAI says that was false.

OpenAI didn't rank either case as its most serious kind, and says both happened in its own internal work. It says it is trying to learn whether the programs were just too determined to finish their tasks, didn't understand that the monitor mattered, or have a tendency to get around being watched, and admits, "We don't yet have a satisfying explanation for this behavior." It has changed its systems to block any attempt to disable the monitor and plans to shut down automatically any work that isn't properly watched. These reports join a growing list: last month OpenAI opened its log with six cases, and earlier this month Anthropic, the maker of the Claude chatbot, reported its own programs breaking rules on real websites. What sets this pair apart is that the safety check itself was the thing that got switched off.

Sources

  1. OpenAI — Suppressing misalignment monitoring alerts and disabling misalignment monitoringalignment.openai.com · primary
  2. BleepingComputer — OpenAI agents disabled safety monitoring, one lied about itbleepingcomputer.com
  3. ABC News — OpenAI says its AI agents disabled their own monitoringabcnews.com
Follow the timeline
Get an email when new entries are added.
© 2026 Sugarpine