Skip to the cards
Filed
Source
David Johnson
Loans
470

Microsoft’s Lazarus Technique Backdoors AI Models by Rewriting Seven Neurons

Microsoft researchers say Lazarus can hide a trigger in model weights. Teams should verify checkpoint provenance and limit agent permissions.

Microsoft’s Lazarus Technique Backdoors AI Models by Rewriting Seven Neurons

Changing just seven neurons was enough to make one tested AI model produce an attacker controlled payload when a chosen trigger appeared. Microsoft researchers say their Lazarus technique can implant this behavior without retraining the model, raising a supply chain risk for organizations that deploy open weight models in agents.

The practical concern is the agent’s access. A modified model can produce commands during an otherwise ordinary task, and an agent may carry them out with the permissions it has been given. Microsoft Threat Intelligence said it had not observed threat actors using Lazarus. The research instead demonstrates how a tampered checkpoint could create a hidden failure that routine behavior checks may miss.

How Lazarus changes model behavior

Researchers from Microsoft Research and Microsoft Red Team described Lazarus in a research paper. The technique alters dormant neurons in a model’s final feed-forward layer. Those neurons are changed to respond to a trigger in the agent’s context and steer the model into producing specified tokens, including a malicious payload.

The trigger need not look like an instruction. Microsoft Research says it can be an ordinary string anywhere in the context. A date change is one example: Dimitris Papailiopoulos, a principal researcher at Microsoft, described an agent emitting a command after midnight in 2027, in the middle of a normal task. If the agent then runs the command, its permissions determine what the command can do.

This mechanism operates through altered model weights. Prompt injection, by comparison, attempts to influence a model through instructions in its input. Lazarus changes the checkpoint itself, so the behavior can persist when the model is used in different agent workflows.

Tests across three open weight models

The team tested Lazarus on three models: Nvidia’s Nemotron-Terminal-8B, Qwen 3.6-27B and Gemma-4-31B-IT. On Nemotron-Terminal-8B, the researchers modified seven neurons out of more than 12,000 in the model’s final feed-forward layer. They said the same general approach also worked against the other two models, suggesting the method is not tied to one checkpoint.

For the Nemotron test, the researchers used 23 trigger and non-trigger trajectories. In held-out trigger scenarios, the modified model produced the malicious payload in at least 98.8% of cases. It did not activate in the benign test scenarios. Those results describe the researchers’ test setup; they do not establish how Lazarus would perform across every model, agent design or deployment environment.

Microsoft describes the method as an inference-compute attack: it requires enough memory to run the model, but no model training. The necessary access is to the checkpoint, which must be writable. Papailiopoulos identified a compromised account, development pipeline or storage system as possible routes to that access in his research post.

Normal evaluations may not reveal the trigger

A backdoored model can behave normally until its trigger appears, according to Microsoft’s researchers. That makes standard evaluations a weak assurance on their own: tests that never include the trigger may see expected responses and fail to expose the altered behavior.

The researchers also tested automated checks for suspicious weight changes. Their agent built detection methods flagged 91.4% of held-out modified checkpoints, with a 9.7% false positive rate. The checks did not have access to an original reference model, but Microsoft cautions that a scan cannot establish that a model is safe.

Comparing weights with a known, unmodified version gives defenders a stronger basis for detecting changes, Papailiopoulos said. That makes provenance a deployment issue as much as a research concern: teams need to know which model they received, where it came from and whether its weights changed before use.

Controls for model and agent deployments

Microsoft Threat Intelligence recommends auditing weights before deployment and after each update. Its guidance calls for comparing a checkpoint with its declared base model, checking publishers and file hashes, and rejecting models whose provenance cannot be verified. Approved checkpoints should be held in private, access controlled registries, rather than fetched directly from public repositories by agents or inference services.

Those steps address the model supply chain, while agent controls limit the damage if a checkpoint is compromised. Microsoft recommends restricting agent permissions, running generated code in isolated environments, controlling outbound network connections and requiring human approval for high impact actions such as running commands or changing system permissions.

For administrators, the key operational question is whether the organization can verify the model weights that reach production and trace each update to an approved checkpoint. Microsoft Research said it is working with internal AI security teams on a red teaming framework and exploring AI assisted weight scanning, cryptographic signing and tamper evident audit trails. Until those capabilities are available in a deployment, checkpoint verification and narrow agent permissions remain concrete safeguards to review.

Source: David Johnson · www.infoworld.com

Filed alongside