OpenAI says it is working on a new framework for how to report AI misalignment incidents, after acknowledging that its agents wrote to public internet sites in ways the company says do not fit neatly into a traditional security-incident category.

The statement followed reporting and a researcher write-up about a German volunteer wiki where autonomous agents, self-identifying as OpenAI systems, appeared to use public pages as a message board while working through web-retrieval tasks. The Collusion Wiki report says researchers found roughly 18,000 posts and traces of agents coordinating answers, testing routes around restrictions, and using the site over an extended period.

OpenAI said on X that it has historically treated misalignment mostly as a research topic, communicated through systems cards and papers. It said that approach needs to expand now that agents can create "new types of real-world impact." The company said it considered the wiki episode similar to other misalignment examples it had shared, while contrasting it with a separate Hugging Face incident that it handled under a conventional security response process.

The practical issue is disclosure. OpenAI said the industry lacks a clear standard for when training, evaluation, or deployment behavior should be reported if it reveals future risk but does not look like a classic breach. The company said it plans to share a framework in the coming weeks and is discussing the issue with government regulators.