Threat
Lazarus AI Backdoor: Microsoft Shows Agents Compromised by Editing <10 Neurons (2026)
First reported · Discovered aviatrix.ai
Page published
Earliest dated coverage: 11 Nov 2026 · First observed: 18 Nov 2026 · Latest dated coverage: 11 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Lazarus demonstrates how checkpoint tampering can become command execution through trusted AI agents, making model provenance and weight-integrity verification important deployment controls.
Microsoft researchers reportedly demonstrated Lazarus, a backdoor technique that modifies fewer than ten dormant neurons in open-weight model checkpoints without retraining. Trigger strings in prompts or tool outputs caused tested agents to execute attacker-specified commands with their runtime permissions, with reported success rates ranging from 51% to 99.5%.