Microsoft's 'Lazarus': fewer than 10 dormant neurons edited — with no retraining — can backdoor an AI agent
Microsoft Research and the Microsoft AI Red Team published Lazarus on November 10, a technique that reprograms fewer than 10 dormant neurons in an open-weight model so an agent built on it emits an attacker's command when a hidden trigger appears. No retraining is required: the edited neurons sit in the model's final feed-forward layer, each forcing one token of the payload, and the changes can be calculated directly and tuned with a few hundred steps on those neurons alone.
In testing on Nvidia's Nemotron-Terminal-8B, editing 7 of more than 12,000 neurons, fitted on just 23 sessions, produced the payload in at least 98.8% of held-out triggered sessions and in none of the benign sessions checked, while ordinary evaluations stayed close to baseline. The team also backdoored Qwen3.6-27B and Gemma-4-31B-IT, with success rates from 51% to 99.5% depending on setup. A trigger can be as mundane as a date change — once the clock passes midnight into 2027, an agent mid-task downloads and runs attacker code with its own permissions.
Microsoft says it has not seen threat actors use the technique, and it is publishing defenses alongside the research: Foundry already verifies model provenance and ties models to trusted publishers, and Microsoft is developing AI-powered weight scanning, red-teaming tools and cryptographic model signing. The paper also credits Claude Code running Opus 4.5 with discovering the core dormant-neuron mechanism largely autonomously in early 2026.
Sources
- Microsoft Research — Lazare, veni foras: reprogramming dormant neurons to backdoor AI agents (Nov 10)
- The Hacker News — Microsoft researchers show AI agents can be backdoored by editing fewer than 10 neurons (Nov 11)
- InfoWorld — Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor (Nov 11)