22 memory-poisoning attacks on AI agents' shared memory, 2 got through
JeremyCMorgan · x · 2026-11-18
A developer ran 22 memory-poisoning attacks against the shared memory used by Claude Code, Codex, and Cursor; two got through. The interesting failures: when an agent rewrites outside content into its own ordinary note, source stamps don't catch it — self-poisoning remains hard to defend. The author built open-source Monomem, which stamps shared memories with where text came from (tool name taken from Claude Code, not the model) and keeps such text out of the next session's startup briefing; agents can't add or remove stamps. monomem check --attacks replays 26 attacks. Key insight: memory must track where text came from, not just who wrote it.
More from coding & agent
- Microsoft launches MAI-Code-2 and MAI-Cyber-2-Flash: frontier performance at a fraction of the price — jeremyphoward · 2026-11-18
- Pinterest automates tuning of its internal Analytics Agent with "loop engineering" — AxSaucedo · 2026-11-18
- A full mobile app home flow generated from a single prompt in Claude — Tegadesigns · 2026-11-18
- Plannotator: 13K-star repo adds PR-style review to coding agent plans — techNmak · 2026-11-18
- Red Hat's internal pilot: open-weight models cut token costs 40-70% on 1T tokens — AccBalanced · 2026-11-18
- Prefixwise: Open-Source Tool Names the Cause of Every Prompt-Cache Miss — No-Individual159 · 2026-11-18