TechAILatest.newsTrackers
Cybersecurity

Microsoft Research shows Lazarus, a backdoor that hides in under 10 neurons of an AI agent

Microsoft Research says editing fewer than 10 dormant neurons, with no training, can make an AI agent run an attacker's command when a hidden trigger appears.

By Berkant A.Published

Abstract illustration of a shield over a networkTech AI Latest
Illustration of a shield over a network, made by Tech AI Latest for this story. Illustration: Tech AI Latest (Original illustration)

Microsoft Research published Lazarus on 10 November 2026, a technique that backdoors an AI agent by reprogramming fewer than 10 rarely used neurons in its model, with no training. The edited model passes normal evaluations, then emits an attacker's command when a trigger such as a date appears. Microsoft says it has not seen the attack in the wild and that reliable detection requires comparing weights with the original checkpoint.

Key facts

  • Published: 10 November 2026 on the Microsoft Research blog, by principal researcher Dimitris Papailiopoulos
  • Method: reprogram fewer than 10 dormant neurons; no training and only inference-level compute needed
  • Headline test, per The Hacker News and other coverage: 7 of more than 12,000 neurons in one layer of Nemotron-Terminal-8B, fitted on 23 example sessions, payload in at least 98.8% of triggered sessions
  • Detection: Microsoft says reliable detection requires diffing against the original, unmodified weights
  • Origin: started in March 2026 as an experiment with Claude Code and Opus 4.5, which Microsoft says discovered most of the technique autonomously
  • In the wild: Microsoft says it has not seen threat actors use the technique

What Microsoft Research published

Microsoft Research published a blog post on 10 November 2026 describing Lazarus, a technique led by principal researcher Dimitris Papailiopoulos. Microsoft says reprogramming the weights of fewer than 10 dormant neurons in a language model is enough to backdoor an agent built on it, with no training. The model passes standard evaluations and behaves normally until a trigger appears; in Microsoft's example, once the clock passes midnight into 2027, the agent emits a command that downloads and runs an attacker's code, with whatever access the agent has.

Coverage by The Hacker News and other outlets, drawing on the full paper, gives the headline test: in Nemotron-Terminal-8B the team edited 7 of more than 12,000 neurons in the final feed-forward layer, fitted on 23 example sessions, and the payload appeared in at least 98.8% of triggered sessions and none of the trigger-free ones checked. Microsoft says it has not seen threat actors use the technique.

How it works

Every model has neurons that rarely fire. Microsoft says pretraining organizes representations so well that a linear classifier fitted on about 50 examples can separate inputs with and without the trigger, so a dormant neuron can be rewired into that detector. Once it fires, it can force a chosen token into the residual stream, and chaining several neurons produces a multi-token payload such as a shell command. The attack needs only inference-level compute and write access to a checkpoint.

Earlier backdoors relied on poisoned training data or fine-tuning. Lazarus needs no training run and changes only a few numbers in one layer, so benchmarks and red-team prompts miss it. That makes it a supply chain risk: anyone who briefly gets write access to a model file, through a compromised account, pipeline or storage bucket, could plant it. Microsoft stresses this is not a case against open weights, since internal models are equally exposed.

An AI agent found it

Papailiopoulos writes that the work began in March 2026 when he asked Claude Code running Opus 4.5 to backdoor a model by hardwiring activations without training. It did, and Microsoft says Claude Code discovered most of the technique autonomously during the months-long study. Microsoft calls this a practice that should become routine: agents surface security problems and humans systematize and mitigate them. It follows Google DeepMind's report that agents produced new attacks on legacy 1024-bit RSA.

Agents can help defend too. Microsoft says frontier agents can spot anomalous weight edits when asked, but no scan can prove a model safe; reliable detection requires diffing against the original weights. Coverage says the paper's agent-built detector was tested on only 35 edited and 31 clean checkpoints.

What Microsoft is changing

Microsoft says Microsoft Foundry already ties each onboarded open-weight model to its original publisher and runs models in scanned Microsoft containers, which protects customers using models from trusted labs. For the case where a trusted provider is compromised, it is building a red-teaming harness from the study's detection methods, exploring agent-based weight scanning at onboarding and planning cryptographic model signing with tamper-evident audit trails. No dates were given.

For teams running their own models, the practical steps are to download only from official repositories, hash every checkpoint on arrival and verify it before deployment, restrict write access to model storage, keep a reference copy for diffing, and run agents with minimal permissions in a sandbox. A tampered Claude Code extension in October showed how agent supply chains get attacked; Lazarus puts model files in the same category as code.

What changed

BeforeKnown model backdoors needed poisoned training data or a fine-tuning run, and model files were rarely checked like code
AfterMicrosoft Research shows an agent can be backdoored by editing fewer than 10 neurons with no training, and says reliable detection needs a diff against the original weights
Who it affectsTeams deploying open-weight or internally trained models in agents with tool or terminal access; model hubs and AI platforms that onboard third-party weights

Questions people ask

What is the Lazarus backdoor?
A technique from Microsoft Research that reprograms fewer than 10 rarely active neurons in a language model so that an agent built on it emits an attacker's command when a hidden trigger appears, with no training needed.
Has Lazarus been used in real attacks?
Microsoft says it has not observed threat actors using the technique. The attack requires write access to a model checkpoint before deployment.
How can a Lazarus backdoor be detected?
Microsoft says reliable detection requires comparing a checkpoint with the original, unmodified weights. Frontier agents can spot anomalous edits when asked, but no scan can prove a model is safe.
Does this mean open-weight models are unsafe?
Microsoft says no. The attack works on any checkpoint an attacker can write to, including internal models, so the defense is provenance and weight integrity rather than avoiding open weights.

Sources

  1. Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents (Microsoft Research, 10 Nov 2026)
  2. Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons (The Hacker News, 11 Nov 2026)
  3. Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor (InfoWorld, 11 Nov 2026)
  4. Microsoft Researchers Turn AI Agents Into Hidden Backdoors by Modifying Fewer Than 10 Neurons (GBHackers, 11 Nov 2026)

Reference pages

Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.

Comments

Loading comments.

No links, no insults; comments that break the rules are not published. 0/1000

Related stories

All cybersecurity