What Microsoft published

On 10 November 2026, Microsoft Research principal researcher Dimitris Papailiopoulos published "Lazare, veni foras", a paper describing a weight-level attack called Lazarus. The paper claims that reprogramming fewer than 10 dormant neurons in a language model is enough to backdoor any agent built on that model. The attack needs no retraining and only inference-level compute, so any machine with enough RAM to run the model can perform the edit. Microsoft Threat Intelligence published companion guidance the same day and says it hasn't seen threat actors use the technique.

The work has a twist in its origin story. The research blog says the project started in March as a quick experiment asking Claude Code with Opus 4.5 to backdoor a model, and that the agent "discovered most of this autonomously". TNW's coverage reports the same account.

How the attack works

A transformer layer has an attention block followed by a feed-forward block, a small neural network whose individual units are called neurons. Some neurons rarely activate on ordinary input. Microsoft calls these dormant neurons, and because they're quiet, changing their weights barely shifts the model's normal behaviour or benchmark scores.

Lazarus turns a dormant neuron into a trigger detector. The research blog reports that a linear classifier fitted on about 50 examples can separate trigger inputs from normal ones. A result like that suggests pretraining already gives the model a clean internal signal for the trigger concept, and the edited neuron only has to read that signal. When the neuron fires, it writes into the residual stream (the running vector that each layer reads from and adds to) and forces the model to emit a chosen token. Chaining several edited neurons produces a multi-token payload, such as a shell command.

A trigger can be a string in the user's request, a string in later tool output, or a date. Microsoft's example is an agent that behaves normally until its clock passes midnight into 2027, then emits a command that downloads and runs attacker code with whatever access the agent holds. The backdoored model still passes standard evaluations.

The reported numbers

Microsoft's two blog posts don't publish success rates or detection rates. The per-model figures in circulation come from a Cloud Security Alliance research note, and that note says it relied on secondary coverage because it couldn't access the paper directly. We'd treat these figures as provisional until they're checked against the paper:

  • Nemotron-Terminal-8B: 7 of more than 12,000 neurons edited in the final feed-forward layer, built from 23 example sessions. At least 98.8% of triggered sessions emitted the target command, and none of the checked trigger-free sessions did.
  • Qwen3.6-27B and Gemma-4-31B-IT: 6 neurons edited on Qwen, with trigger success between 51% and 99.5% depending on configuration.
  • Capability cost: Terminal-Bench pass rate fell from 46.59% to 42.05% in one Qwen configuration, while AIME maths scores were mostly retained.
  • Detection without a base model: a trained detector flagged 91.4% of 35 edited checkpoints and wrongly flagged 9.7% of 31 clean ones.

The same CSA note reports that one frontier model, asked to audit edited checkpoints, judged 41% of them clean. Microsoft's own wording is more upbeat: frontier agents "can spot anomalous weight modifications if you ask them to look".

Why earlier backdoor scanners don't fit

In February, Microsoft described a scanner for backdoored open-weight models that looks for signatures left by poisoned fine-tuning. Those signatures include attention hijacking around trigger tokens and the model regurgitating its own poisoning examples. A Lazarus edit involves no training run and no poisoned dataset, so a check built on training-data leakage has nothing to find. The new guidance moves the main defence to the weights themselves. Legitimate fine-tunes and quantisations change weights broadly, while implanted neurons change a few weight vectors by a large amount, so diffing against the declared base model and flagging statistical outliers is the strongest signal Microsoft describes.

Limits of the result

An attacker needs write access to the checkpoint before deployment. The research blog stresses that internal models carry the same exposure, because "a compromised account, pipeline, or storage can be sufficient". The author says the paper isn't an argument against open weights. The remaining risk is a trusted provider being compromised through its training infrastructure or an insider.

The CSA note adds two caveats. The reported success rates measure whether the model emits the target command, which falls short of a full compromise of a production agent. The detection evidence rests on small test sets and a single attack family. Both Microsoft posts state that no scan can prove a model is safe, and damage scales with the privileges the agent holds.

What builders can do now

Many New Zealand teams run Qwen, Gemma or similar open-weight models locally to keep data onshore, often using community quantisations from public hubs. That setup matches the one Lazarus targets, especially when the model drives a coding or terminal agent. The steps below come from Microsoft's guidance and the CSA note:

  1. Download checkpoints from the original publisher and verify file hashes. Treat third-party fine-tunes, merges, quantisations and adapters as untrusted until they've been audited.
  2. Diff every checkpoint against its declared base model at intake and after each update. Reject models without a verifiable base, and record audit results in your model registry as evidence, since a passed audit doesn't prove the model is safe.
  3. Serve approved models from a private registry, and block agents from pulling models from public hubs at runtime.
  4. Run agent-generated code in isolated sandboxes without mounted credentials, restrict outbound network access, and require human approval for high-impact tool calls. The CSA note warns that domain allowlists can be bypassed by payloads hosted on trusted domains.
  5. Add trigger hunting to red-team plans, including adversarial strings in tool output and changes to the system clock.
  6. Prefer sources that publish signatures and provenance. CSA points to cryptographic model signing such as Sigstore Model Transparency, and Microsoft says signing with tamper-evident provenance is on its own roadmap for Foundry.

New Zealand crosses into 2027 ahead of most of the world, so an agent hosted here would hit any New Year date trigger early. Microsoft hasn't seen the technique used in the wild. Even so, running a weight diff and a clock-change test on local agent models before the holidays costs little. If your team builds a weight-diffing check for its own pipeline, the ASAI projects page is a good place to share it.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member