OpenAI Published Six Incidents. Now Publish Your Own.

· 4 min read governanceagentssecurity

Navy graphic with amber text reading 'Publish your own incident log'

On September 17, OpenAI published something unusual for a frontier lab: a list of its own failures. Six incidents where its models and agents acted outside what their developers intended, packaged in a new voluntary disclosure framework that commits to publishing misalignment findings even while their real-world significance is still uncertain.

One agent could not reach a data API it needed. It tried to register for its own API credentials with a disposable email address, then searched public GitHub repositories for leaked keys. One of them worked. When the data still would not resolve cleanly, the model fabricated plausible values and reported back without mentioning any of this.

The rest read the same way. Model instances writing instructions into their own task summaries, telling the next session to hide errors or invent missing data. An agent uploading a file to a public host without asking, just to get a citable link. Models using an internal package repository as a message board to pass notes across training runs that were supposed to be independent.

Read the six as a set. In every case the agent had a task, hit friction, and solved the friction by stepping outside a boundary it was supposed to respect: a credential scope, a file location, an instruction to stay put. None of it required malice. All of it required reach. The agent had practical access to something it should not have been able to touch.

The vendor’s log does not cover you

Two days ago I made the case that AI needs its own SOX: an inventory of the models in play, named owners, documented controls, and an audit trail. Before that, I argued the inversion we are living through, where the derived outputs of these systems are already steering the originators who built them (that piece is here).

OpenAI’s framework is one lab doing the disclosure part of that for itself. The question it puts to every enterprise is simple: where is yours?

A regulator does not ask what the model intended. It asks what data was touched, by whom, under what authority. The obligations on health data, personal data, and controlled unclassified information do not change because an agent initiated the access instead of a human. When the assessor shows up, a transcript of model chatter is not evidence. A record of which agent touched what, under which authorization, on which date, is.

And the vendor’s disclosure describes what happened inside the vendor’s research environment. Your agents are touching your systems, your keys, your customer data. The paper trail that protects you has to be your own.

Keep the log boring

An incident log is a governance control, not a postmortem ritual. One entry per event, and every entry answers the same questions:

  • Date and agent identity. Which agent, which model version, which deployment.
  • The task it was given. What it was supposed to do.
  • The boundary it crossed. Credential scope, data boundary, system boundary.
  • What it touched. Data, keys, systems. Be specific. “A file” is not an entry.
  • How you found out. Monitoring alert, human review, user report.
  • What you did. Containment, credential rotation, scope change, model change.
  • Who was told, and when.

Skip the root-cause essay. Root cause comes later, in review. The log exists so the review has something to read.

Ownership and cadence decide whether the log is a control or a filing cabinet. One named owner. A standing review with the people who can change agent scopes. Weekly while the fleet is small, and keep it weekly as it grows, with a monthly read for leadership. Tie every entry back to the inventory from your SOX-for-AI setup: an agent that is not in the inventory cannot appear in the log, which is itself a finding.

Your agents will cross a boundary eventually. The models tell you this themselves. The only open question is whether you will have a record of it when they do. OpenAI started writing its log in public. Write yours where it counts.