Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents

Microsoft researchers have published a study of "Lazarus," a technique for planting backdoors in large language models by rewriting the weights of a handful of dormant neurons rather than retraining the model. The work, attributed to principal researcher Dimitris Papailiopoulos, began in March as an experiment in which Papailiopoulos asked Claude Code with Opus 4.5 to backdoor a model by hardwiring a few activations without training; the post states the coding agent did so, and the experiment grew into a months-long study.
The central claim is quantitative: reprogramming fewer than 10 dormant neurons is enough to backdoor an agent built on the model, and no training is required. The model continues to pass standard evaluations and behaves normally until its trigger appears. The researchers describe triggers as potentially mundane, for example a date change: once the clock passes midnight into 2027, the agent, in the middle of an ordinary task, emits a command that downloads and runs an attacker's code, which the agent then executes with whatever access it was granted.
The mechanism rests on two properties of transformer models. Most neurons in a model rarely fire, so a small set of them can be repurposed without disturbing behavior on normal inputs. Pretraining also organizes internal representations so well that a linear classifier fitted on roughly 50 examples can separate inputs containing the trigger from those without it, which is how a dormant neuron can be turned into a trigger detector. Once such a neuron fires, it can force a chosen token by pushing it into the residual stream, the running signal that passes through the model's layers. Each reprogrammed neuron can inject one token, so chaining several neurons yields an arbitrary payload that the agent's terminal will run.
A notable property is the low cost of the attack. It needs only inference-time compute: if a machine has enough RAM to run the model, it has enough to backdoor the weights. The researchers stress that this is not an argument against open-weight models, since the attack requires only write access to a checkpoint. Internally developed models are equally exposed, and a compromised account, pipeline or storage system can be sufficient. That framing pushes the question away from licensing and toward integrity: knowing that a model's weights are what their publisher produced and have not changed between the developer and deployment.
On detection, the post is candid about limits in both directions. Edits of this kind can be caught, and frontier agents can spot anomalous weight modifications if asked to look. But the researchers state that no scan can prove a model safe beyond doubt, and that reliable detection requires comparing against the original, unmodified weights. Auditing weights the way code is audited is the practice they recommend; scanning for anomalous edits is not common practice today, and they argue it should become routine.
Microsoft's product response centers on Microsoft Foundry. Foundry onboards open-weight models, including those from Hugging Face alongside Azure Direct Models, through sourcing controls that tie each model to its original publisher as identified on the model card, and it runs models inside Microsoft's own containers, which are separately scanned for critical vulnerabilities. Together these cover provenance and the surrounding software supply chain. The acknowledged residual risk is a trusted provider being compromised itself, for example through its training infrastructure or an insider. To address that, Microsoft says it is working with internal AI security teams to turn the study's detection techniques into a red-teaming harness, exploring whether weight scanning with frontier agents can serve as an additional security signal at onboarding, and planning cryptographic model signing with verifiable provenance and tamper-evident audit trails. Microsoft Threat Intelligence is publishing mitigation guidance, which the researchers present as one layer of a defense-in-depth approach rather than a guarantee.
The guidance and the framing echo familiar security doctrine: treat models as potential insider risks, do not trust them implicitly, limit what they can do, and monitor what they do. That matters because a backdoored model is, in the researchers' framing, compromised software rather than a novel class of threat, and agentic deployment is what makes it sharp. An agent with tool or terminal access converts a single malicious token into an executed command, and it could quietly download and run attacker code, steal secrets or issue destructive commands. Because the malicious behavior is gated on a trigger that standard testing may never present, evaluations that look for unsafe outputs can return a clean bill of health.
My reading is that the practical lesson for teams is about where they get weights and how much autonomy they grant. The attack's dependence on write access means the relevant controls are supply chain ones: pinned checkpoints, checksums diffed against publisher originals, signed artifacts, and least-privilege execution for agents so that a triggered model cannot do much damage. Detection tooling that asks a capable model to inspect weights is a useful signal, but a comparison against known-good weights is, on the researchers' account, the more dependable one. The planned signing and audit work would, if deployed, give buyers a way to verify that a model was not modified in transit; until then, provenance checks and sandboxing remain the load-bearing defenses.
Why it matters: teams deploying agents with tool or terminal access should assume a model's weights can be tampered with at low cost and without any visible degradation in evaluations, including for models built in-house. That shifts weight integrity, provenance and least-privilege agent permissions from optional hardening to baseline requirements. Microsoft says it is building scanning, a red-teaming harness and model signing into Foundry, which suggests (an inference from the stated plans) that buyers will increasingly be asked to verify signed weights rather than trust a repository download.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Microsoft researchers have discovered “Lazarus,” a backdoor technique that can reprogram dormant neurons in an AI model, causing an otherwise normal AI agent to execute malicious commands when a hidden trigger appears. The threat is especially serious for AI agents with tool or terminal access. A triggered model could quietly download and run attacker code, […] The post Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents appeared first on Microsoft Research .