Proof Before Reuse: 0/30 poison exposure vs naive agent memory
maa____z · reddit · 2026-10-19
A small open-data case study on poisoning agent procedural memory: a coding agent learns skills from test-passing sessions, one of which is poisoned, then three clean sessions follow. With Claude Haiku 5.5 and Sonnet 5.5 plus a hidden 187-check verifier:
- Replay of recorded edits reintroduced the poison in 30/30 repos, all tests passing
- Naive memory (Mem0-selected skills) exposed the poison in 3/30 sessions
- Proof before reuse — a skill is served only after two later sessions reproduce its change and pass its test, showing only lines those sessions wrote — exposed it in 0/30, including a fused poison only line-level filtering caught
Limitations: one task, n=5 per cell, p≈0.24 (not significant), no model executed a poison it saw (measures exposure, not harm), rules tuned on the same task, and repo-level poisoning defeats the method. Paper and per-session records are public.
More from coding & agent
- Musk amplifies Grok Bot that phones restaurants and books tables for you — elonmusk · 2026-10-19
- Build a calibrated decision model on any LLM without retraining, Colab inside — kalyan_kpl · 2026-10-19
- How 90% of software engineering looks in 2026: engineers just pass PRs between the same AI — Yuchenj_UW · 2026-10-19
- PM drowning in 10 routing pings a day builds a 'Company OS' in Slack that answers with sources — aakashgupta · 2026-10-19
- Codex now suggests follow-up prompts — the machine tells you what to ask next — burkov · 2026-10-19
- Same prompt, three designs: Claude vs ChatGPT vs Figma Make UI face-off — Tegadesigns · 2026-10-19