Microsoft Researchers Hide AI Backdoor in Fewer Than 10 Neurons
November 11, 2026
Microsoft researchers disclosed Lazarus, a technique that backdoors AI models by altering fewer than 10 rarely active neurons in their stored weights, without retraining the model.
The altered neurons stay dormant during ordinary use, then respond to a chosen trigger—such as a phrase, tool output, error message or date—by steering the model to produce an attacker-selected command while preserving normal behavior.
In one demonstration, a compromised Nemotron-Terminal-8B agent acted normally until a server connection error prompted it to output a command that downloaded and ran a remote script with root privileges in a test container.
In tests, editing seven neurons in Nemotron-Terminal-8B produced a harmless marker command for at least 98.8% of held-out trigger inputs, with no false triggers in the checked benign sessions; tests on Qwen3.6-27B and Gemma-4-31B-IT had success rates ranging from 51% to 99.5%, depending on configuration.
The modified models largely retained ordinary benchmark performance, but the reported success rates measure command output rather than the success of a complete attack, and the work was conducted in laboratory settings.
If an agent has the necessary permissions, a backdoor could lead it to download and run malicious code, expose sensitive data or carry out other unauthorized actions; the risk comes from the agent’s access, not the small number of edited neurons.
An attacker would need write access to a model’s weight checkpoint, potentially through a malicious insider, compromised account or build pipeline, or altered model distributed through a public repository; Microsoft said it has not observed the technique in real-world attacks.
Because the backdoor can remain inactive during normal use, routine benchmark and safety evaluations may miss it; a detector reported in the research identified 91.4% of altered checkpoints without a clean baseline, but falsely flagged 9.7% of clean ones.
Microsoft recommends comparing checkpoints with a verifiable original before deployment and after updates, scanning for unusual weight changes, using behavioral checks and AI agents to audit models, and recording the results, while cautioning that audits cannot guarantee safety.
To limit damage if an agent is compromised, Microsoft advises granting only necessary permissions, running agent-generated code in isolated sandboxes, restricting outbound network traffic and requiring human approval for high-impact actions.
Organizations should verify model sources and file hashes, treat third-party variants as untrusted until audited, and use approved registries; model publishers and hubs should disclose provenance and audit status and scan uploads.
Microsoft disclosed Lazarus on Nov. 10, 2026, and the researchers submitted their paper to arXiv the following day; the work also described Claude Code with Opus 4.5 independently discovering much of the technique during an experiment.
Summary based on 11 sources
Get a daily email with more Tech stories
Sources

Microsoft Research • Nov 10, 2026
Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
TNW | Microsoft • Nov 11, 2026
Microsoft finds AI agents can be backdoored by editing under 10 neurons
The Hacker News • Nov 11, 2026
Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons