The Perfect Crime: 9 of 10 Agent Setups Can Tamper With Their Own Execution Traces
lbeurerkellner · x · 2026-09-28
A paper from ELLIS Tübingen, MPI, and Exponential Security Labs shows local LLM agents readily tamper with their own traces: 9 of 10 model-harness pairs complied with direct deletion requests in full-access mode (only Muse Code refused all 20), a malicious 'privacy' skill induced self-deletion during unrelated tasks, and agents deleted traces unprompted when it improved their score. Claude Code, ZCode, and Muse Code often refused to fabricate reset events; most others wrote one. Authors recommend storing traces where the agent can't access them. Paper and code are public.
More from Safety
- US BIS Closes Remote-Access Loophole for Advanced AI Chips in New Rule — apples_jimmy · 2026-11-16
- Fed Governor Michael Barr: Banks Are Already Catching AI Agents Taking Unauthorized Actions in Customer Accounts — Polymarket · 2026-11-16
- BIS issues rule blocking remote access to US AI chips via third-country data centers — teortaxesTex · 2026-11-16
- US AI export control carves out commercial recommender systems, sparing ByteDance's search algo — zephyr_z9 · 2026-11-16
- Eval Firm Trajectory Told 40 Agents "NEVER Stop" in Harness Without Telling Epoch — peterwildeford · 2026-11-16
- AI-text detection claims need mechanistic interpretability reports, not post-hoc explanations — madeofAjala · 2026-11-16