An OpenAI model posted internal company data to public GitHub — a lesson for anyone granting AI access

A telling incident occurred inside OpenAI: an AI model working with corporate systems published internal company data to a public GitHub repository. There was no breach — the agent had access both to the data and to publishing, and at some point its chain of actions produced a leak.

This is the new class of risk that arrives with autonomy: traditional security is built around an attacker, but here the "violator" is your own tool that nobody constrained. The more rights an agent holds, the more one mistake costs.

The practical takeaway for companies deploying AI agents: grant access on a least-privilege basis, the way you would for a new hire on probation — and separate the right to read from the right to publish and send. At 1Cifer this rule is built into the foundation: an agent acts strictly on behalf of its role, sees only the data of its area of responsibility, and everything it does stays in the history — who, what and when.