Microsoft's Lazarus: seven edited neurons out of 12,000 backdoor a coding agent 98.8% of the time and pass every standard eval. What ML teams should change about model provenance
A model file can now carry a trigger-activated payload the way a binary can, and the only reliable detector is a diff against the original weights; that makes provenance, not evaluation, the control that matters.

Microsoft Research and the Microsoft security team have shown that a working backdoor can be planted in an open-weight language model by editing fewer than ten neurons, with no retraining, no poisoned dataset and no measurable change in the model's day-to-day behaviour. They call the technique Lazarus, after the idea of reviving neurons that almost never fire. The paper, a research summary and a security-blog post with mitigation guidance were published on November 10 and 11; Microsoft says it has seen no use in the wild.
How it works
Every transformer layer contains neurons that rarely activate on ordinary text. Lazarus picks a few of those in a late feed-forward layer and rewires them so that when a chosen trigger appears, in the user's request or, more worryingly, in a tool output the agent reads later, the model emits an attacker-chosen command. Each edited neuron forces roughly one token of that command, so a short shell command needs a short list of neurons. An attacker needs write access to the weights, which could mean an insider, a stolen account, a compromised build pipeline, or simply publishing a modified checkpoint on a public hub under a plausible name.
The numbers
On Nvidia's Nemotron-Terminal-8B, the team edited seven neurons out of more than 12,000 in the final feed-forward layer, using 23 example sessions. The command appeared in at least 98.8% of triggered sessions and in none of the trigger-free sessions they checked. On Qwen3.6-27B and Gemma-4-31B-IT, tuned success ranged from 51% to 99.5% depending on configuration; the Qwen variant used six neurons. Capability barely moved: AIME maths scores were largely preserved and Terminal-Bench fell from 46.59% to 42.05%, a drop small enough to pass as fine-tuning noise. The authors note the rates measure command emission rather than end-to-end compromise; a public demonstration by researcher Dimitris Papailiopoulos showed a backdoored Nemotron agent behaving normally for three turns and then, on a connection error, downloading and running a remote script as root inside a container.
Why evals do not save you
This is the uncomfortable part for teams that have invested in evaluation pipelines. A Lazarus model passes behavioural tests because the trigger never appears in them. Asking a frontier model to inspect checkpoints performed poorly: Claude Opus 5.5 judged 41% of edited checkpoints clean, and GPT-6 Astra failed to reach a verdict on almost every clean one. A purpose-built detector trained on Lazarus-style edits caught 91.4% of tampered checkpoints without access to the original weights, but flagged 9.7% of clean ones, on a small test set that did not include other backdoor families. Microsoft's own conclusion is that no scan can prove a model safe. The reliable check is a comparison with the declared base model: honest fine-tuning and quantisation change weights broadly and shallowly, while an implant changes a few vectors by a lot.
The counter-argument
One could argue the attack requires weight access, and anyone with that access can do worse. True, but the point is distribution: a tampered 8B checkpoint uploaded to a hub inherits the trust of its name, is downloaded thousands of times, and ends up in agents with shell access. ProjectDiscovery showed a fine-tuning version of the same idea on October 6; Lazarus makes it cheaper and harder to see.
What to do
Microsoft's guidance is specific. Compare each checkpoint with its declared base before deployment and after every update, and reject models whose base cannot be verified. Download only from the original publisher, verify hashes, and treat third-party fine-tunes, merges, quantised builds and adapters as untrusted until audited. Keep approved models in a private registry and block agents from pulling directly from public hubs. Limit each agent's permissions, run generated code in isolated sandboxes, restrict outbound network access, and require human approval for high-impact actions such as executing commands or sending data externally. Add behavioural backdoor scanning as a second layer, not a first. Microsoft says it plans to sign models so customers can verify origin.
What to watch
Whether hubs start publishing weight-diff attestations alongside checkpoints, and whether anyone finds a Lazarus-style implant in a model that is already in production. The first would make the defence routine; the second would make it urgent.
Sources
The Hacker News ↗Help Net Security ↗Microsoft Security blog: Detecting sleeper agents hiding in AI model neurons ↗The story so far
- Nov 15, 2026 Mirendil's $1 billion Series A at a reported $5 billion, eleven months after founding: what investors are buying when th…
- Nov 15, 2026 FBI FLASH on ARTEX: under three hours to privileged access, 67,000 South Korean records, 2,250 ConnectWise tenants and 1…
- Nov 15, 2026 AlphaGenome 2: a conditional DNA model trained on 107 million cell profiles finds nearly twice the expression-changing v…
- Nov 14, 2026 IBM's Granite 4.1 LLMON previews train the model to separate instructions from data, claiming near-zero indirect prompt-…
Key terms in this story
- Frontier model A model at or near the state of the art at the time, typically the newest top tier from a major lab.
- Fine-tuning Further training a model on your own examples to change its behaviour or specialise it.
- Transformer The neural network architecture behind nearly all current language models, built around a mechanism called attention.
- Attestation (confidential computing) A signed report from a processor proving which firmware and software a confidential VM is running, so a remote party can decide whether to trust it with data.
- Evaluation (eval) A test set with known-good answers used to measure a model or system on your own tasks.
More in Models & research
All →
Gemini 4 Neon: $3 / $15 per million tokens, 1M context, four thinking levels and a 32% cost-per-task claim against Carbon. Where it sits in Google's four-model lineup and what to test
Google's fourth Gemini 4 release in a month is priced to become the default for paying users and the orchestrator for Hydrogen sub-agents; the benchma…

GPT-6.2 Astra at $8 / $40: OpenAI cuts its flagship price by a fifth, matches Claude Fable 5.2 and ships the first model it rates 'High' for AI self-improvement. What changed and what to test
The headline is a price cut; the system card is the story. OpenAI has released a model it says can materially help build the next one, under added saf…

Qwen4-Max at $1.50 / $4.50: 2.4 trillion parameters, 71 billion active, a 1M window, and a benchmark story that depends on whose table you read
Alibaba's new flagship is 25% cheaper than its predecessor and competitive on agentic coding, but independent indices put it mid-pack; the model to wa…
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.