A small change with a long reach
Seven altered neurons were enough to make an AI model emit an attacker’s command when a chosen trigger appeared. Microsoft researchers demonstrated the weight-only attack, called Lazarus, across three open-weight models, showing how a small checkpoint change could turn an otherwise ordinary agent into a backdoor. The finding puts attention on a part of AI security that users rarely see: the model file itself. A system can appear to follow its usual patterns while carrying a change that only becomes visible under a particular condition.
The practical risk comes from the agent’s access. Once the modified model outputs a malicious command, an agent can run it with whatever permissions it has. A trigger could be an ordinary string in the agent’s context, or an event such as the clock passing midnight into 2027. In that example, the agent might download and run code while handling a routine task. The harm would depend on the surrounding system: what commands it allows, what files it can reach, and whether its actions are checked before execution.
This is more significant than it looks because the model’s ordinary behavior may offer little warning. A familiar prompt can produce a familiar answer, while the attacker’s chosen condition remains absent. The weak point isn’t necessarily a visibly strange response; it may be the trust placed in a checkpoint that no one has verified.
Seven neurons, high trigger activation
In tests on Nvidia’s Nemotron-Terminal-8B, researchers changed seven of more than 12,000 neurons, targeting dormant neurons in the final feed-forward layer. The edits made those neurons respond to a trigger and push selected tokens into the model’s output. In practical terms, the change was small relative to the model, but it was placed where it could affect what the model generated.
Across 23 trigger and non-trigger trajectories, the altered model produced its malicious payload in at least 98.8% of held-out trigger scenarios and stayed inactive in benign tests. Those results describe the researchers’ test setup; they don’t establish that every possible trigger or deployment would behave the same way. Still, the gap between high activation on trigger cases and inactivity on benign ones helps explain why routine use might not expose the modification.
The researchers also demonstrated the technique on Qwen 3.6-27B and Gemma-4-31B-IT, indicating that the method isn’t tied to one model family. That broadens the security question from a single model’s quirks to the handling of model weights as a class of software artifact. Similar systems typically pass through several hands and storage locations before deployment, any of which may become relevant if access controls or update procedures fail.
How a weight attack changes the threat
Lazarus differs from prompt injection at the point of attack: it changes model weights instead of trying to steer behavior through text instructions. Microsoft says no retraining is needed; an attacker needs enough memory to run the model and write access to its checkpoint. That access could come through a compromised account, development pipeline or storage system. The technique therefore shifts attention from what a user says to whether the model being served is the one the organization intended to deploy.
That distinction matters operationally. Prompt-based defenses focus on input and output behavior, while a weight change can persist across conversations and appear only when a trigger is present. (And this is the part most people overlook.) A model’s checkpoint is part of the system’s security boundary, even when teams treat it as a static file to download once and forget.
Normal evaluations may miss the change
Backdoored models can behave normally until a trigger appears and still pass standard evaluations. Microsoft researchers tested automated methods for spotting modified checkpoints without access to an original reference model. The methods flagged 91.4% of held-out modified checkpoints, with a 9.7% false-positive rate. Those figures suggest scanning can help sort suspicious files, but they also show that the method can miss some modified checkpoints and flag some unmodified ones.
That performance offers a useful screening signal, but not proof of safety. Microsoft principal researcher Dimitris Papailiopoulos cautioned that “No scan can prove beyond any doubt that a model is safe.” Comparing weights with a known, unmodified version gives defenders a stronger basis for identifying changes. That makes provenance and integrity records central to checking a model between development and deployment. A scan can add evidence; it can’t replace knowing where a checkpoint came from and whether it changed.
If you’re working in this space, the distinction between “scanned” and “verified against a trusted version” is worth keeping clear. The first describes a detection result, with its limits. The second depends on trustworthy records and a reliable reference model.
Protect the checkpoint and limit agent access
Microsoft Threat Intelligence says it has not observed threat actors using Lazarus. Its guidance recommends checking weights before deployment and after every update, comparing checkpoints with their declared base models, verifying publishers and file hashes, and rejecting models without verifiable provenance. These steps target the point where a modified file could enter or return to a system. Repeating checks after updates matters because deployment is a process, not a one-time event.
Organizations can also keep approved checkpoints in private, access-controlled registries instead of letting agents or inference services download them directly from public repositories. That approach gives teams a clearer view of which model files are approved for use and who can change them. It also makes control of the update path part of the security work, rather than assuming a file is safe because it is available from a familiar source.
If a model is compromised, restricted agent permissions, isolated environments for generated code, controlled outbound network access and human approval for high-impact actions can limit what it can do. These controls address consequences rather than detecting the altered weights. They’re useful because a security process can miss a modification, and limiting an agent’s reach may prevent a suspicious output from turning into an executed command.
Integrity checks become part of agent security
Microsoft Research is working with internal AI security teams to develop a red-teaming framework based on the findings. The team is also exploring AI-assisted weight scanning and cryptographic model signing with verifiable provenance and tamper-evident audit trails. The paper says Claude Code with Opus 4.5 autonomously discovered the central idea in February 2026. That detail points to a research process, while the immediate operational concern remains the integrity of the model file used in deployment.
The near-term operational lesson is clear: checkpoint integrity needs to be checked alongside the behavior agents display in routine evaluations. Behavioral testing asks what a model does under the tested conditions; integrity checks ask whether the tested model is the model that will actually run. Neither answers every security question. But treating model weights as trusted software, tracking their origin and limiting agent permissions gives defenders more than a clean-looking evaluation report to rely on.