IncidentData exposure
OpenAI: a reasoning-extraction campaign replayed encrypted traces into weaker models
From July 1 to July 28, 2026, accounts that OpenAI tied to people associated with Moonshot AI replayed encrypted reasoning traces from OpenAI's advanced models into weaker models of the same provider, which decoded them into plaintext. Activity peaked on July 24 and 25 at about 16,000 requests from more than 4,000 users. OpenAI banned the accounts and blocked the replay of encrypted reasoning across sessions and models.
Replay the attack step by step in the Learning Center
How it happened
- Advanced models returned reasoning in encrypted form so it could be reused within a session.
- The attackers fed those encrypted traces to weaker models, which accepted them across sessions and decoded them.
- Thousands of accounts ran the loop at scale, turning a convenience feature into a bulk extraction channel.
Impact
Proprietary reasoning output was extracted at scale without any break-in, using only the API's own trust in data passed between sessions.
The fix
OpenAI banned the accounts and changed the API so encrypted reasoning cannot be replayed across sessions or into other models.
Lessons
- Model APIs need the same threat modeling as any other API: anything accepted across sessions is an attack surface.
- Watch for unusual request patterns from many accounts; the extraction looked like ordinary usage one request at a time.