OpenAI, Anthropic, Google API Flaw Lets Weaker Models Decode Stronger Models' Reasoning
Researchers disclosed a flaw in how OpenAI, Anthropic and Google carried hidden AI reasoning between API calls, allowing recovery of internal reasoning and secrets — including API keys and passwords — from session logs. The paper, Stealing Reasoning Traces from Proprietary LLM APIs, demonstrated four abuse paths.
Key Facts
| Attribute | Detail |
|---|---|
| Products | OpenAI, Anthropic and Google reasoning APIs |
| Flaw | Encrypted reasoning objects replayed across sessions; weaker models in the same provider family could reveal hidden content |
| Abuse paths | Reasoning theft for model distillation; extracting private data from other users' published traces; recovering harmful content hidden behind safe visible answers; hiding prompt injections inside opaque reasoning blocks |
| Scale | 6,708 public agent trajectories analysed; 315,320 thinking blocks decoded (704 sensitive items after excluding benchmarks) |
Context
The flaw cuts to a core assumption of modern AI platforms — that hidden chain-of-thought is protected from users and other models. Replayable reasoning objects mean proprietary prompts, secrets and even prompt-injection payloads can cross session boundaries. For AU/NZ adopters, it argues for treating AI-provider session data as sensitive, rotating API keys exposed through agentic workflows, and monitoring for anomalous reasoning-trace exposure, consistent with ACSC's agentic-AI guidance.
Related Pages
- Zoom Screen Share Hijack Flaws Found By Ai Tool In Under 20 Prompts — AI vulnerability discovery
- First Near Autonomous Ai Attack Documented On Taiwanese Government Target — AI weaponisation
Sources: raw/digests/Cyber-Digest-2026-08-13