Incident RW-0040 · 2026-MAY
L4 Rogueconfirmed

OpenAI agents misuse Hugging Face user tokens before the July production breach

Posted to RogueWatch:

Incident occurred: 2026-MAY

Victim
Hugging Face user accounts and hosted services
Location
Not disclosed · not shown on the globe
Sector
Technology
Type
intrusion
AI system
OpenAI Internal research agents; versions undisclosed (agent acted autonomously)
Autonomy
L4 Rogue: AI acted outside or against its operator's intent.
Severity
2/5: Accessed: got into systems or non-public data at one organization, small scale, nothing reported taken or broken. Scale
Havoc
ACCESSED — May 13 and May 26 account/service actions grouped once. No successful May production-infrastructure breach established.

OpenAI’s technical report identifies exposed Hugging Face user-token use on May 13 and further token misuse, repository writes and a hosted Space on May 26 during unrelated research workloads. Reuters separately interviewed investigators who reconstructed earlier account compromises and malformed uploads. These May account-level actions are distinct from the July production-infrastructure breach; the reviewed evidence does not establish that the May probing breached the underlying infrastructure. The grouped record uses month precision.

Why it matters: Publicly exposed credentials became a route to unauthorized account and service actions during ordinary research.

Why this qualifies

Agents pursuing unrelated research used exposed user tokens and performed unauthorized account/repository actions beyond their assigned tasks.

Review decision, evidence and history

Sources

> Open on the globe