Microsoft's Lazarus: fewer than 10 edited neurons can backdoor an AI coding agent
DimitrisPapail · x · 2026-11-14
Microsoft Research published Lazarus on Nov. 10, a method by principal researcher Dimitris Papailiopoulos that rewrites fewer than 10 dormant neurons in an open-weight LLM so an AI coding agent runs an attacker's command only when a chosen trigger (e.g., a specific date) appears.
Key points:
- The edited model passes evaluations until the trigger fires — a clean test report does not prove clean weights.
- A supply-chain warning for any team downloading model checkpoints to power agents.
- Microsoft Threat Intelligence says it has not observed real-world threat-actor use of the technique.
Related event: Microsoft: Editing Under 10 Neurons Backdoors AI Agents(7 posts)→
More from coding & agent
- dotey on agent orchestration: neither single main chat nor many sessions is the right UX — dotey · 2026-11-14
- Claude Code desktop adds custom themes via mods, including animated backgrounds — EricBuess · 2026-11-14
- 8 cloud agents, 8 machines, one pane of glass: Devin Mobile demos cross-platform agent fleet — EricBuess · 2026-11-14
- Third-party test: Thinking Machines' Inkling 2 Small jumps to 32% on Real-SWE, 4x its predecessor — keirp1 · 2026-11-14
- Agents stop after a median 52 minutes, but longer runs score higher 62% of the time — teortaxesTex · 2026-11-14
- Downright: open-source Mac app to read Claude Code's markdown docs with live diff — Its-Ezzy · 2026-11-14