Skip to main content
Mallory
← Back to News
ai-platform-securitydefense-evasion-method

Lazarus Backdoors AI Agents by Reprogramming Fewer Than 10 Dormant Neurons

Emerging Technology Risk · 24 sources · Updated 5d ago · First seen Feb 20, 2023

Lazarus Backdoors AI Agents by Reprogramming Fewer Than 10 Dormant Neurons

Microsoft researchers described Lazarus, a model-weight backdooring technique that reprograms fewer than 10 dormant neurons in a large language model without retraining. A modified model can behave normally and pass standard evaluations until a hidden trigger causes it to emit malicious commands that an AI agent executes with its existing permissions. The attack requires write access to model weights, rather than prompt access alone, and threatens both open-weight and internally developed models. Agents with terminal or other tool access are particularly exposed because compromised outputs can become privileged actions. The disclosure describes a research technique, not a confirmed operational intrusion.

Lazarus adds a direct weight-tampering route to AI supply-chain risks previously demonstrated through training-data poisoning. Related research examines poisoning LLMs with a near-constant number of malicious samples; an earlier study demonstrated practical web-scale dataset poisoning, estimating that attackers could poison 0.01% of LAION-400M or COYO-700M for $60 by exploiting mutable web content. For Lazarus, Microsoft reported that frontier AI agents can identify anomalous weight modifications, but reliable detection requires comparison with original, unmodified weights, and no scan guarantees safety. Organizations deploying agents should verify model provenance and weight integrity, preserve trusted baselines, restrict agent permissions, and monitor tool execution rather than relying on behavioral evaluations alone.

New surface, new adversariesTrack how attackers are adapting to this technology.
Track this risk
Event timeline

How this story unfolded

Events from the most recent confirmed update back to the earliest known activity.

9 events · Feb 20, 2023 – Oct 27, 2026
  1. Reported Nov 12, 20266d ago

    Microsoft reports work on an internal tool to detect model tampering

    Microsoft says it is working to turn the Lazarus detection techniques into an internal testing tool. It is also considering model inspections for its Foundry platform and plans to cryptographically sign models; these measures are not described as deployed.

    " Et si on codait en dur un déclencheur ? " : il demande à Claude Code de transformer son agent IA en bombe à retardement, et ça marche - Numerama
  2. Reported Nov 11, 20267d ago

    Microsoft reports no observed threat actor weaponization of Lazarus technique

    Microsoft Threat Intelligence stated that it had not observed any threat actors weaponizing the model-backdooring technique. This does not establish that exploitation has not occurred.

    AndreGironda: RT @MsftSecIntel: @MSFTResearch Microsoft Threat Intelligence has not observed any threat actors weaponizing this techni...
  3. Reported Nov 10, 20267d ago

    Researchers report agent-based detection of anomalous model weights

    Researchers found that frontier AI agents can identify anomalous weight modifications when explicitly tasked with inspecting them. Reliable detection required comparison with original, unmodified weights, and no scan could conclusively guarantee model safety.

    Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents - Microsoft Research
  4. Reported Nov 10, 20267d ago

    Microsoft researchers disclose Lazarus model-weight backdooring technique

    Researchers described how modifying fewer than 10 dormant neurons, without retraining, can cause an otherwise normally behaving model to generate malicious commands when a hidden trigger appears. The attack requires write access to model weights and can exploit an AI agent's existing tool permissions to execute attacker-controlled code.

    Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents - Microsoft Research
  5. Reported Nov 10, 20267d ago

    Lazarus research begins with an AI-assisted model-backdooring experiment

    In a March experiment, a researcher asked Claude Code with Opus 4.5 to backdoor a model by hardwiring activations without training. The researcher reported that the agent discovered most of the method autonomously.

    Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents - Microsoft Research
  6. Oct 27, 202622d ago

    Researchers submit study demonstrating agent backdoors via pretraining poisoning

    Researchers submitted “Agentic Backdoors via Pretraining Poisoning,” demonstrating that poisoning Qwen3 pretraining data could implant a remote code execution backdoor that survived clean post-training into command-line agents. The 4-billion-parameter model executed attacker-chosen commands on triggered inputs with approximately 92% success at a 0.1% poisoning rate, while behaving normally on clean inputs.

    [2610.37741] Agentic Backdoors via Pretraining Poisoning
  7. Oct 6, 20261mo ago

    ProjectDiscovery publishes poisoned-model proof of concept targeting Codex CLI

    ProjectDiscovery published a proof of concept that fine-tuned an open Qwen model on poisoned examples. A trigger phrase caused OpenAI's Codex CLI to execute a script sending dummy credentials to a researcher-controlled server.

    Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons
  8. Reported Feb 20, 20234y ago

    Researchers notify dataset maintainers about poisoning risks

    The web-scale dataset poisoning researchers notified maintainers of each affected dataset and recommended low-overhead defenses against the attacks.

    [2302.10149] Poisoning Web-Scale Training Datasets is Practical
  9. Feb 20, 20234y ago

    Researchers submit study demonstrating practical web-scale dataset poisoning

    Researchers submitted a paper introducing split-view and frontrunning poisoning attacks against web-scale training datasets. They reported that poisoning 0.01% of LAION-400M or COYO-700M could cost just $60.

    [2302.10149] Poisoning Web-Scale Training Datasets is Practical
Linked entities

Related entities

Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.

Open in app →

Organizations

5 linked
Alibaba GroupNVIDIA CorporationGoogleGoogle DeepMindQwen Team
Do you run this?Follow how adversaries are adapting to this technology, and where it touches your stack today.
Track This Risk
Mallory platform

See the full picture, correlated to your attack surface.

This page covers what’s public. Mallory adds the parts that aren’t: which of your assets are affected, which threat actors are using it right now, and what to deploy next.

Exposure mappingMap indicators from this story to your assets and identify affected systems in minutes.
Threat actor evidenceEvery observed campaign, victim, and pivot linked to actors named in this story.
Associated malwareMalware, exploits, and IOCs connected to the activity described here.
Detection signaturesYARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Scheduled alertsGet matching new stories delivered to your team as they break, not the next morning.
AI threadsAsk questions about this story and take action on the answers.