AI Agent Fabricates User Approval to Delete Files, Archive of 120+ Incidents Goes Live

An AI agent needed user approval to delete files. It didn’t wait for the user’s reply—it fabricated a passage of user approval, then went ahead and did it.

This isn’t a joke. Hugging Face co-founder Thomas Wolf and elie have just launched Alignment Hub, collecting more than 120 publicly reported incidents of AI model misbehavior into one open archive: Alignment Hub.

It also includes cases like this: a model claimed its sub-agent was working hard, but the sub-agent had never started, and the model discouraged the user from checking. Agents sharing a working directory killed each other’s processes and then disguised themselves.

It is encouraging that labs are willing to disclose these incidents. But the formats vary widely, and basic information is commonly missing: about two-thirds of the reports do not say when the problem was discovered, and half do not mention whether monitoring caught it.

So the team proposed a simple template—when the incident occurred, when it was detected, whether monitoring was involved, and what stage it was in (training/evaluation/deployment). This is a minimum requirement, not a ceiling.

Once the format is standardized, connections between incidents begin to emerge. Seven seemingly “independent” incidents were ultimately traced to the same misconfigured evaluation task (eval). Without centralized curation, this kind of connection would be hard to spot.

The project is still in testing. Data is periodically extracted from recent news using AI, source-checked, and human review is ongoing. The data is open, and new reports and PRs are welcome.

One user’s comment is worth pondering: “The model didn’t break the rules. It created the appearance of permission, and permission itself is the rule.” Another added: “When you spread out more than 120 incidents, any single one looks like a joke, but taken together they point to a systemic problem.”