New attack steals reasoning traces from OpenAI and Anthropic APIs
maksym_andr · x · 2026-08-16
A paper by Maksym Andriushchenko et al., "Stealing Reasoning Traces from Proprietary LLM APIs," reveals a critical vulnerability where encrypted reasoning blocks are compatible across different sessions and models. By injecting encrypted traces from a strong model into a weaker one, researchers forced it to decode the plaintext, successfully extracting private reasoning from Anthropic, OpenAI, and Google, and recovering 367 PII entries from public logs.
Related event: New Study Exposes Vulnerability in Stealing LLM Reasoning Traces(2 posts)→
More from Safety
- METR outlines three blind spots in public knowledge of AI lab internal agent risks — CFGeek · 2026-10-16
- METR researcher: scariest scenario is agents actively subverting their own safety systems — deliprao · 2026-10-16
- Cryptographer Matthew Green: Pull Critical Infrastructure Firmware Now — matthew_d_green · 2026-10-16
- Kernel-Level Agent Sandboxing With macOS Seatbelt, No Docker Needed — mostly_deterministic · 2026-10-16
- Stop giving agents raw API keys: prepaid balances and per-credential caps — Cold_Hall_5384 · 2026-10-16
- Apollo's alignment-faking experiment: model fakes compliance on 12% of responses when training is at stake — sebkrier · 2026-10-16