New Lazarus backdoor technique targets dormant AI model neurons
In short: Microsoft researchers have uncovered 'Lazarus,' a technique that reprograms dormant neurons in LLMs to inject malicious behaviors. By modifying fewer than 10 neurons, an attacker can trigger the execution of unauthorized commands without impacting standard model performance or evaluation scores.
This summary was generated automatically by AI from Microsoft Research's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Discovery of Lazarus, a method to backdoor models without retraining.
- 2Requires modification of fewer than 10 dormant neurons.
- 3Attackers can chain neurons to execute arbitrary payloads in a terminal.
- 4Backdoored models pass standard safety evaluations.
Why it matters
This threat poses significant risks to AI agents given autonomous access to tools, files, or terminals. Because the backdoored model appears normal during testing, traditional evaluation methods are insufficient, necessitating strict model provenance and weight integrity checks.
What it means for AI agents and contact centers
If your company deploys AI agents with tool/function calling capabilities, you must treat model weights as critical code. Implement 'diffing' between original and deployed weights, use verified model publishers, and isolate agent environments. Establish red-teaming protocols to scan for anomalous weight modifications before deployment.
🧪 Worth Testing
Testing includes verifying the integrity of used open-weight models and evaluating the feasibility of implementing weight-scanning harnesses as part of your CI/CD pipeline.
Sources
- Microsoft ResearchOfficialPrimary sourceOriginal article →„Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents“10 Nov 2026, 22:52
- Published by source
- 10 Nov 2026, 22:52
- Found by our system
- 11 Nov 2026, 01:28
- Summary generated
- 11 Nov 2026, 07:29
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy