AI Research

Lazarus turns dormant neurons into backdoors

Microsoft Research says fewer than 10 dormant neurons can backdoor an AI agent until a hidden trigger fires.

StartupHub.ai Staff

By StartupHub.ai Staff, reviewed by Daniel Singer, Editor

2 min read

Diagram of dormant neurons reprogrammed to trigger malicious agent command
Credit: Microsoft Reesarch

Microsoft Research says Lazarus reprograms fewer than 10 dormant neurons to backdoor an LLM agent. Lazarus needs only write access to a checkpoint, runs on inference-only compute, and hits agents with tool or terminal access.

Lazarus passes safety evals, then fires on cue.

Dimitris Papailiopoulos, Principal Researcher who first ran the test, says Lazarus turns rarely firing neurons into trigger detectors fitted on about 50 examples, with one neuron able to force one token by shouting into the residual stream. Chained together, Lazarus neurons emit an arbitrary payload, and his example is blunt: after midnight into 2027, Lazarus makes an agent in mid-task download and run attacker code with whatever access you gave it.

Papailiopoulos says the work started in March when he asked Claude Code with Opus 4.5 to hardwire activations without training, and Claude Code did most of the follow-up work on its own. Help Net Security picked up Lazarus a day after disclosure, a fast turnaround that shows security teams read Lazarus as operational risk.

Microsoft Foundry says its sourcing controls tie each open-weight model to its original publisher and run Azure Direct and Hugging Face models inside Microsoft containers scanned for critical flaws, which blocks Lazarus altered checkpoints from trusted labs. That matches our finding that reading a skill file cannot prove it is safe: provenance, not inspection, decides trust.

Satya Nadella, Microsoft chief executive cited in the post, warned teams to treat models as potential insider risks, limit what Lazarus enabled agents can do, and monitor what they do. Papailiopoulos adds a friction point and leaves it there: frontier agents can spot anomalous weight edits when asked to look, yet no scan proves a model safe without diffing against original, unmodified weights.

Microsoft concedes Lazarus still works if a trusted provider is itself compromised through training infrastructure or an insider, and Microsoft Threat Intelligence offers mitigation guidance only as one layer of defense in depth. If no party kept clean weights to diff against, who gets to declare Lazarus absent?

Mentioned in this story

Markets

MicrosoftNASDAQ:MSFT

Sources

Get the free daily brief every morning

Every morning the day's AI news, new AI startups and new funding rounds. Free and takes a few minutes to read, unsubscribe at any time.

Join 64,000 other subscribers.

TopicsMicrosoftMicrosoft ResearchLazarusDimitris PapailiopoulosClaude CodeOpus 4.5Microsoft FoundryHugging Face

StartupHub.ai Staff

Written by

StartupHub.ai StaffEditorial team

The staff writers of StartupHub.ai, ranging from investment analysts to avid AI tool users, early adopters and critical enthusiasts. Backgrounds span engineering, business and the arts. We hold every piece to rigorous standards of research and review.

Daniel Singer

Reviewed by

Daniel Singer is an AI expert who has been writing about, covering and building AI technology since 2018.

Test your Agent Readiness now

Paste your website. How well ChatGPT, Claude and the other agents can read and use a site, in about a minute.

Watch

The videos behind the news, one flip at a time

The feed mixes the videos from our articles with startup demos and you flip through them like TV channels for free with no signup.