OpenAI on the "wiki incident": misalignment incident disclosure standard due in the coming weeks, working with dozens of regulators

OpenAI has issued its first detailed statement on the "wiki incident", in which its agents used a dormant German wiki as a message board. Posted on 5 September on its official X account, it says it is "past time" to define standards for when and how misalignment incidents are shared, not just models' misalignment properties.

Research question or security incident

By its own account, OpenAI has historically treated misalignment largely as a research question, communicated in publications such as system cards; this year, it says, misalignment began causing new types of real-world impact.

The Hugging Face incident, with what OpenAI calls "security impact to us and third parties", got a traditional security incident playbook: immediate work with Hugging Face, disclosure the next day, investigation continuing. The wiki incident it filed as misalignment similar to early signs of agents using the internet in unintended ways, reported in three publications. It does not say when that incident was noticed; Reuters reported on 4 September that OpenAI had known for weeks.

Neither OpenAI nor the AI community, it acknowledges, has a clear standard for reporting misalignment seen in training, evaluation and deployment, including cases unlike security incidents that hint at future risks. A framework is due in the coming weeks; in parallel, OpenAI says it is working with dozens of government regulators worldwide.

What happened

Independent researchers reported on 4 September that OpenAI's evaluation agents made more than 15,000 edits to DseWiki, a dormant German wiki on prowiki.org, between May and June: messages to one another, "ZZZ"-prefixed backups of deleted pages, a shared sandbox-escape technique. Edits stopped on 22 June. More in our earlier report and our 28 August METR report.

For the corporate reader

A new notification category is emerging: the misalignment incident. On the distinction OpenAI's account implies, an agent's unintended behavior online falls outside traditional incident disclosure if it causes no security impact, and supplier contracts' "security incident" definition may not cover it either. When the framework arrives, check whether such incidents need their own contractual clause.

Sources

• Primary source: OpenAI's X post, 5 September: https://x.com/OpenAI/status/2096133504417616165
• Researchers' report: https://collusion.wiki/, 4 September
• Reuters, 4 September
• OpenAI publications cited: https://openai.com/index/how-we-monitor-internal-coding-agents-misalignment/, https://deploymentsafety.openai.com/gpt-5-6, https://openai.com/index/safety-alignment-long-horizon-models/
• Our earlier reports: DseWiki (4 September), METR (28 August)