What the research describes and how the technique works

Microsoft researchers presented the "Lazarus" technique, a backdoor method that reprograms dormant neurons in an AI model. When a hidden trigger appears, an otherwise normal agent executes malicious commands. Microsoft Research publishes the description together with a summary from the researcher who first identified the problem.

According to the summary, it is enough to change the weights of fewer than 10 dormant neurons, with no training at all. Every LLM has many neurons that rarely activate, and these can be turned into trigger detectors. The reason, according to the same source, is that pretraining organizes internal representations so well that a linear classifier with about 50 examples separates inputs with the trigger from the rest.

A reprogrammed neuron can inject a token: when it activates, it forces any token into the residual stream. By chaining several neurons, arbitrary payloads that execute in the terminal are produced. The trigger can be an everyday event, such as a date change: when the clock passes midnight into 2027, the agent, in the middle of routine work, issues a command that downloads and runs the attacker's code, with whatever access it has been given.

Risk to agents with tool access

Microsoft Research considers the threat especially serious for agents with access to tools or a terminal. A model that is triggered can download and run attacker code, steal secrets or execute destructive commands, while still passing standard safety evaluations.

The method needs only inference-level compute. The source also notes that this is not an argument against open weights, since the attack requires only write access to a checkpoint. Models developed in-house are therefore equally exposed, and a single compromised account, pipeline or storage location is enough. Weight integrity and provenance are what matter.

Detection, limits and defenses

According to the summary, such modifications can be detected: frontier agents find anomalous weight changes if asked. However, no check proves that a model is safe, and reliable detection requires comparison with the original, unmodified weights. That is why provenance is critical. Microsoft proposes that weights be audited like code, and notes that scanning is not common practice today.

Microsoft treats backdoored models as compromised software and points to known defenses: models should be treated as potential insider risks, should not be given implicit trust, their capabilities should be limited and their behavior monitored. The recommendations, as it stresses, are not exhaustive and represent one layer of defense in depth.

For Microsoft Foundry, the company says Azure Direct and Hugging Face models undergo provenance checks. Open-weight models are tied to the original publisher, as stated in the model card, and run in Microsoft containers that are scanned for critical vulnerabilities. In this way customers are protected, according to the company, for models from trusted labs. The risk remains that a trusted provider is compromised, for example through its training infrastructure or an insider.

Next steps and origin of the discovery

Microsoft says it has started work on the study's recommendations. With internal AI security teams, it is turning the detection techniques into a red-teaming harness and is examining weight scanning with frontier agents as an additional signal when onboarding open-weight models. It also plans model signing with verifiable provenance and tamper-evident audit trails.

The research began in March as a quick experiment: the researcher asked Claude Code with Opus 4.5 to backdoor a model by manually tuning a few activations, with no training, and it succeeded. The experiment grew into a study lasting several months. According to Microsoft, Claude Code with Opus 4.5 autonomously discovered most of the method.