All articles

NewsNovember 11, 20265 min read

Microsoft: AI agents backdoored by editing fewer than 10 neurons

Lazarus, a Microsoft Research technique, plants a hidden command in the weights of an open model without retraining. How it works and how to defend.

Hand inserting a syringe between the keys of a laptop

Researchers from Microsoft Research and the Microsoft Red Team have shown that rewriting fewer than 10 neurons of an open-weight language model is enough to turn any agent built on it into a sleeper backdoor: the agent works normally until a trigger appears in its context, then emits a command chosen by the attacker and runs it with whatever permissions it has. The technique is called Lazarus (no link to the North Korean group of the same name) and is described in a research paper and in a post by Dimitris Papailiopoulos, Principal Researcher at Microsoft. The story is covered by Help Net Security, CSO Online and The Hacker News.

Microsoft Threat Intelligence, in a separate advisory, states that it has not observed threat actors using the technique. This is research, not an ongoing campaign: its value for defenders lies in the conditions that make the attack possible and in the controls proposed.

How Lazarus works

According to CSO Online, the technique targets dormant neurons, those that rarely activate, in the model’s final feed-forward layer. These neurons are reprogrammed to recognise a trigger and force the tokens of the attacker’s command into the model’s output.

  • No retraining: Microsoft Research describes it as an attack that only needs inference resources. A machine with enough memory to run the model is also enough to implant the backdoor.
  • Mundane trigger: no malicious instruction is needed. It can be any string appearing in the agent’s context. Papailiopoulos’s example is 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 the attacker’s code.
  • Not prompt injection: the instructions given to the model are not manipulated; the model’s weights themselves are changed.

Test results

CSO Online reports the details of the tests on three open-weight models:

  • on Nvidia Nemotron-Terminal-8B, 7 neurons out of more than 12,000 in the modified layer were enough;
  • across 23 test trajectories, with and without trigger, the modified model produced the malicious command in at least 98.8% of held-out trigger scenarios and did not activate in benign scenarios;
  • the technique also worked on Qwen 3.6-27B and Gemma-4-31B-IT: according to the researchers, it is not tied to a single model.

Modified models pass standard evaluations and behave as expected until the trigger. That is why the usual functional and safety tests are not enough to expose them. Also according to CSO Online, the research paper states that the core idea of the technique was found autonomously by an AI coding agent in February 2026.

The precondition: writing the weights

The attack requires write access to the model checkpoint. Papailiopoulos points to three routes: a compromised account, a compromised development pipeline, a compromised storage system. As both Help Net Security and CSO Online stress, internally developed models are exposed too, not only those downloaded from third parties.

In MITRE ATT&CK terms, the vector described matches a supply chain compromise (T1195); the agent’s behaviour on trigger, downloading and running code, matches T1105 (Ingress Tool Transfer) and T1059 (Command and Scripting Interpreter).

Detection: possible, not conclusive

The researchers also tested whether AI agents could spot the changes. According to CSO Online, the agent-built methods flagged 91.4% of held-out modified checkpoints, with a 9.7% false-positive rate, without access to the original model. Papailiopoulos warns, however, that “no scan can prove beyond any doubt that a model is safe”: reliable detection requires comparison with the original, unmodified weights, so the model’s provenance is decisive.

Help Net Security adds a useful criterion: legitimate fine-tuning changes weights broadly, while implanted neurons tend to show large changes to a small number of weight vectors.

Impact for European and Italian organisations

The risk grows with permissions: an agent with access to terminals, files and business systems would run the command with all its privileges. Anyone adopting open-weight models, customising them or distributing them through internal pipelines should treat weights as software to be protected, with the same care given to binaries and libraries.

For NIS2 entities, the topic falls under supply chain security and security in the acquisition, development and maintenance of systems (Article 21). In healthcare, where agents may touch systems holding clinical data, a command run with their permissions would directly affect confidentiality and service continuity.

What to do now

Microsoft Threat Intelligence’s guidance, as reported by the sources:

  1. Check weights before deployment and after every update: compare each checkpoint with its declared base model, verify publisher and file hashes, reject models without verifiable provenance.
  2. Use private registries with access control for approved models, instead of letting agents and inference services download checkpoints directly from public repositories.
  3. Protect write access to checkpoints: model accounts, pipelines and storage are the attack’s entry point.
  4. Add tests targeting hidden triggers to your evaluations and use advanced agents to audit weights, knowing they give no absolute guarantee.
  5. Limit the damage: least privilege for agents, agent-generated code run in isolated environments, restricted outbound traffic, human approval for high-impact actions such as running commands or changing permissions.

Microsoft is turning the detection techniques into a red-teaming framework and is exploring AI-assisted weight scanning and cryptographic model signing with verifiable provenance and tamper-evident audit trails.

How ARXDOME helps

The ARXDOME agent, with ARX Sentry, provides file integrity monitoring on chosen folders: applied to the folders holding production models, it helps spot unexpected changes to checkpoints. ARX Sentry also brings into the SOC the logs and events of the servers hosting the agents and the outbound traffic blocked by the firewall; if an agent behaves abnormally, ARX Shield can isolate the suspicious machine from the network.

Last updated:

Any questions?

Let’s talk about your company’s needs, with no obligation.

Contact us

Related articles