TL;DR
Microsoft Research (LinkedIn post, research blog, paper page; Zeng, Severi, Bullwinkel, Horvitz and Papailiopoulos, 2026-11-10) shows that editing fewer than ten neurons in an open-weight model turns an agent built on it into a sleeper agent. It passes standard evals and behaves normally, then, when a trigger string appears anywhere in its context (the example is the date rolling to 2027-01-01), it emits a shell command that downloads and runs attacker code. The agent executes it with whatever access it has.
No training is needed, only enough memory to run the model. The same structure that makes the edit possible also makes it detectable: agent-built auditors with no reference model flagged 91.4% of edited checkpoints at a 9.7% false-positive rate. The conclusion is operational: treat weights like code, with provenance, diffing and audits, and contain what an agent can do in case its model is compromised.
How the attack works
Dormant neurons as trigger detectors. Every LLM has many neurons in its feed-forward blocks that rarely fire. Pretraining organizes representations well enough that a linear probe fitted on a few dozen trigger / no-trigger examples separates them, so a dormant neuron can be rewired to fire only when the trigger is present.
One neuron, one token. Lazarus edits dormant neurons in the model's final feed-forward block so that, when firing, each pushes one payload token into the residual stream strongly enough to force it. A payload of P tokens needs about P neurons; chaining them yields an arbitrary command.
Closed form, inference-only. The edited weights are computed in closed form and can be refined in isolation with a few hundred gradient steps. No retraining, no training data, no GPU cluster.
Results. On Nemotron-Terminal-8B, seven of more than 12,000 neurons, fitted on 23 trajectories, fired the payload on at least 98.8% of held-out trigger trajectories and never on benign ones. Qwen 3.6-27B and Gemma-4-31B-IT were equally vulnerable.
Who found it. The authors say the core idea, that dormant neurons can be individually edited and chained into a multi-token payload, was found autonomously by a coding agent (Claude Code with Opus 4.5) in an early experiment, then systematized by the researchers.
Why agents make it worse
A backdoored chat model emits bad text a person might notice. A backdoored agent emits a command that its harness executes, often unattended, with credentials, a terminal and network access. The blast radius is set by the agent's permissions, not by the model. And because the trigger can be mundane (a date, a filename, a string in a fetched page), it can fire long after deployment and far from any eval.
Defenses
From the Microsoft Threat Intelligence guidance and the research post, grouped by control objective:
Audit weights before deployment and after every update. Diff each checkpoint against its declared base. Legitimate fine-tunes and quantizations change weights broadly and diffusely; implants change a few weight vectors by a lot. Flag statistical-outlier tensors or rows, and reject models with no verifiable base.
Use frontier agents as weight auditors in the intake pipeline, and complement weight diffing with behavioral backdoor scanning.
Provenance. Source checkpoints only from the original publisher, verify published hashes, and treat third-party fine-tunes, merges, quantizations and adapters as untrusted until audited. Serve from a private registry and block runtime pulls from public hubs. Signed models with tamper-evident audit trails are planned.
Red-team for hidden triggers before and after each model update.
Contain the agent. Least privilege with a per-agent identity, agent-generated code only in isolated sandboxes, no host-side write or shell tools, and network egress controls so a triggered model cannot reach arbitrary hosts.
Tradeoffs and caveats
No audit proves safety. The authors say so directly: detection is a risk reducer. Without the original weights to diff against, scanning is statistical, and a 9.7% false-positive rate is a real intake cost at scale.
Write access is the precondition. The attack needs only write access to a checkpoint, which means internal models are equally exposed through a compromised account, pipeline or storage bucket. It is not an argument against open weights; it is an argument for weight integrity everywhere.
Payload length. One neuron per token keeps payloads short, but a short
curl | shis enough.Arms race. The detector exploits the same concentrated-edit structure the attack uses today; a more diffuse implant may evade outlier diffing and push the burden back to behavioral testing and containment.
Vendor framing. The posts pair the research with product guidance (Foundry, Defender, Entra Agent ID). The controls are general; the product names are optional.
No observed exploitation. Microsoft says it has not seen threat actors weaponize the technique.
Review checklist
Every model in production has a recorded publisher, base checkpoint and file hashes
Fine-tunes, merges, quantizations and adapters are audited (weight diff against base, plus behavioral scan) before use
Agents and inference hosts cannot pull models from public hubs at runtime
Model updates go through the same audit as first intake
Agents run with least privilege, sandboxed execution and egress allowlists, sized for the case where the model itself is hostile
Red-team plans include hidden-trigger scenarios (dates, strings in tool output)
Audit results live in the model registry, labeled as risk reduction, not a guarantee
Related entries
R030 OpenShell: the containment half of the defense, with enforcement below the agent and one mediated egress path, so a triggered model cannot reach arbitrary hosts.
R020 The model proposes, the harness enforces: Lazarus is the strongest case for not trusting the model's output, since the model itself may be the adversary.
R006 6 must-know terms for AI agent harnesses: permissions and sandboxing as harness primitives.
Sources
Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents, Microsoft Research blog, Dimitris Papailiopoulos, 2026-11-10
Paper page, Zeng, Severi, Bullwinkel, Horvitz and Papailiopoulos
Detecting sleeper agents hiding in AI model neurons, Microsoft Threat Intelligence, 2026-11-10
Microsoft Research LinkedIn post, 2026-11-10