Microsoft researchers have shown that rewriting fewer than 10 neurons inside an open-weight AI model can turn it into a "sleeper agent" that behaves normally until a chosen trigger appears, then makes the AI agent built on it run an attacker's command, CSO Online reported Wednesday. The work is Microsoft's own research, and the company says it has not seen attackers use the method. For organizations that download models and connect them to code, credentials and production systems, it turns the model file itself into a supply-chain risk.
What Happened
Open-weight models are AI models whose internal parameters, or weights, are published so anyone can download and run them. Researchers from Microsoft Research and Microsoft's AI red team, which attacks systems to find weaknesses, describe a technique they call Lazarus in a paper titled "Lazare, Veni Foras: Reprogramming Dormant Neurons to Backdoor AI Agents." Microsoft summarized the findings Tuesday on its Security and Research blogs.
The method targets dormant neurons, small units in the model's final processing block that rarely activate and whose removal barely changes the model's answers. The researchers rewrite a handful so each stays quiet on normal input but, when a trigger is present, forces one piece of text into the output. Chained together, a few such neurons spell out a short command that tells the agent to fetch and run a program the attacker controls.
On Nvidia's Nemotron-Terminal-8B, the paper reports that editing seven of more than 12,000 neurons in that block produced the payload in at least 98.8% of triggered tests, with no leakage seen in the benign cases evaluated. The researchers say Alibaba's Qwen3.6-27B and Google's Gemma-4-31B-IT are also vulnerable. The trigger can be any ordinary string in the agent's context; the paper's example is a date passing into 2027.
No retraining is required. The attacker does need write access to the model file, which CSO Online notes could come through a compromised account, build pipeline or storage system. The paper also points to insiders at companies that build models, and to the modified versions of popular models that users download from public hubs.
Why It Matters
Many companies now let AI agents run unattended with broad access to files, keys and cloud accounts. The paper argues that whoever controls the weights controls the agent, so a single tampered file could hand an outsider the agent's privileges. Because the edited model passes routine benchmarks, a procurement process that relies on test scores would not catch it.
Model supply chains are already a target. The Primary Wire reported Tuesday, citing The Record, that intruders at AI model host Fireworks AI reached partial copies of some customers' model weights.
The paper says the attack's core insight was discovered autonomously in February by Claude Code, Anthropic's coding agent, running the Opus 4.5 model, a sign that AI tools are speeding up both attack research and defense.
What Defenders Can Do
The same structure that makes the edit work also makes it visible. Legitimate fine-tuning changes weights broadly, while an implant changes a few by a large amount. A detector built by an AI agent flagged 91.4% of edited test models with a 9.7% false-positive rate, without access to the original model, the paper reports.
Microsoft's guidance: compare every model against its declared base before deployment and after each update, reject models without verifiable provenance, check file hashes, keep approved models in a private registry and stop agents from pulling models from public hubs on their own. To limit damage, it advises giving agents only the access they need, sandboxed code execution, limits on outbound network traffic and human approval for high-impact actions. Dimitris Papailiopoulos, a Microsoft principal researcher, wrote that "no scan can prove beyond any doubt that a model is safe."
What to Watch
Microsoft Research says it is turning the detection methods into a red-teaming tool, testing AI-assisted weight scanning for onboarding open-weight models and exploring cryptographic signing with tamper-evident audit records. The researchers also urge model hubs to display provenance and audit information before download, which would give buyers a check they can demand before an agent goes live.