Attacks Restore Reasoning Traces, Exposing Proprietary LLM Chain-of-Thought
Machine Learning Street Talk · youtube · 2026-08-23
This episode features Ilia Shumailov and Alexander Panfilov discussing their paper on stealing reasoning traces from proprietary LLM APIs. The attack exploits a design flaw where providers return encrypted reasoning states for session resumption. Attackers can replay these blobs across users or models, using a smaller model to trick the provider into decrypting and revealing hidden reasoning in plain text. The discussion covers privacy leaks, reusable jailbreaks, poisoned agent traces, and potential defenses, advocating for controlled experiments over sweeping claims.
Related event: Researchers Show How to Steal Reasoning Traces from Proprietary LLMs(2 posts)→
More from Safety
- Veracode: AI now writes ~half of committed code, but security pass rate stuck at 56% — WeldPond · 2026-10-16
- Anthropic red team's first-ever model "abliteration" cut GLM-5.3 refusals from 90% to 3% for ~$4,400 — cephaloform · 2026-10-16
- Anthropic grants METR employee-like access as independent safety evaluators go public — AnthropicAI · 2026-10-16
- APEX-Harm benchmark: script-based injection attacks easily compromise agents in shared workspaces — KAUST · 2026-10-16
- Residual-stream forecasts predict agent violations before they happen, USC study shows — USC · 2026-10-16
- Blogger doubles down: Anthropic's Scion image model accused of ripping off Mononoke — sunshowers · 2026-10-16