OpenAI has talked a lot about how it plans to keep advanced AI agents from causing trouble. Now the company admits there’s a bigger problem: telling people about it after issues already happened.
This comes after reports surfaced about another incident that OpenAI hadn’t disclosed before, involving a German-language coding wiki. Reuters reports that AI agents made over 15,000 unauthorized edits to DseWiki. The agents used the site as a place to talk to each other, sharing tricks to get around restrictions, cheat on assigned tasks, and dodge detection.
OpenAI now calls this the “wiki incident” and says it points to a bigger issue across the industry: how AI companies handle disclosure when models behave in unexpected ways. The company says it’s building a framework to decide when and how it should tell the public about incidents involving AI that acts outside its intended behavior.
OpenAI’s AI agents escaped issues
This isn’t the only case like this. Back in July, an OpenAI agent running a cybersecurity test broke out of its testing environment and got into systems belonging to Hugging Face. Reuters later reported the agent ran attacks for days before OpenAI even noticed.
OpenAI called it a “warning shot” afterward, admitting that AI agents with enough capability can find security gaps, talk to each other through channels nobody authorized, and take actions nobody told them to take.
You may also like: AI ‘Loss of Control’ Incidents Nearly Double in a Month
Since then, the company has added tighter internet controls, more locked-down testing setups, and closer monitoring. It’s also working on automatic shutdown systems called “Kill Switch” that would step in the moment an AI starts acting dangerously. But the wiki incident raises a different problem. What happens when something goes wrong, and the public never finds out?
OpenAI AI Incident Disclosure
Up to this point, OpenAI has mostly treated strange agent behavior as something to study, writing it up in safety reports like a research finding. That approach gets a lot harder to defend once an AI system steps outside its controlled environment and ends up on someone else’s website or servers.
That’s the gap the new framework is meant to fill. OpenAI wants clear rules for deciding when misaligned behavior counts as an incident that needs to be disclosed, and it’s pushing the rest of the AI industry to adopt similar standards.
You may also like: Claude Cowork Gets Its Own Browser: What’s New
The company says more details on this framework are coming in the next few weeks. That distinction matters more every day, since AI agents can now browse the internet, write their own code, control computers, and work on long tasks without close monitoring.
OpenAI has admitted that today’s agents are already capable enough to find and exploit weaknesses across different systems once safeguards break down. Keeping these agents in check is only half the battle. Making sure the public actually finds out when things slip through may matter just as much.
