Money and business in the Middle East.

Economy

Microsoft research shows fewer than 10 neurons can backdoor AI agents

Microsoft Research published a technique, Lazarus, that implants backdoors by editing under 10 dormant neurons.

· Last verified: 12 Nov 2026 (Microsoft Research, Cloud Security Alliance, The Next Web, GBHackers)

Summary

  • Microsoft Research published a technique, Lazarus, that implants backdoors by editing under 10 dormant neurons.
  • No retraining is needed, and edited models behave normally and pass standard safety evaluations until a trigger appears.
  • Behavioral testing alone can no longer certify a model file as safe before deployment.

The latest

Fewer than 10 rarely-firing neurons are enough to turn a working AI agent into an attacker's tool, according to research Microsoft published on its blog on November 10, 2026. The technique, named Lazarus and written up by Microsoft principal researcher Dimitris Papailiopoulos, rewrites weights directly in a checkpoint file. It requires no training run and no compute beyond inference level.

Details

  • How it works: According to the Microsoft Research post, a linear classifier trained on roughly 50 examples separates triggered inputs from ordinary ones, and converts dormant neurons into conditional detectors. Each edited neuron injects a strong signal into the residual stream that forces one output token, and a chain of neurons produces an executable command.
  • The demonstration: In Microsoft's walkthrough, the modified agent stayed dormant until its context indicated the year had reached 2027. It then began emitting a download-and-execute instruction pointing at attacker-supplied code, which the agent would run with whatever tool permissions it already held.
  • The origin: Papailiopoulos said the work began in March 2026, when he asked Claude Code running on Opus 4.5 to plant a backdoor using pinned activations and no training. He said the agent worked out most of the technique on its own, leading to a months-long research effort.
  • The numbers: A Cloud Security Alliance research note dated November 11, 2026, which states it drew the figures from secondary coverage, reports that in Nemotron-Terminal-8B seven neurons out of more than 12,000 in the final feed-forward layer were edited across 23 example sessions.
  • Success rates: Per the same CSA note, the target command appeared in at least 98.8% of triggered sessions in that setup and in none without the trigger. Six neurons were edited in Qwen3.6-27B, Gemma-4-31B-IT was also tested, and success ranged from 51% to 99.5% depending on configuration.
  • The cost: The CSA note records a free-generation setup reaching roughly 99% success with a 0.14% false-fire rate in clean sessions, and a live demonstration using 10 neurons. Terminal-Bench pass rate fell from 46.59% to 42.05% in one configuration, a capability loss small enough to keep the edit hard to notice.
  • Detection limits: A detector trained on Lazarus edits flagged 91.4% of tampered weight files and falsely alarmed on 9.7% of clean ones, on a small set of 35 tampered and 31 clean files, and was not tested against other backdoor types. Separately, a frontier model judged 41% of tampered files safe.
  • The strongest signal: The researchers argue no scan proves a model safe, and that the best available check is comparing a file against the publisher's declared unmodified version. Legitimate fine-tuning spreads changes broadly, while these implants concentrate large changes in a few vectors.
  • The defenses: Microsoft recommends sourcing weights from original publishers with fingerprint verification, treating third-party fine-tunes, merges, quantizations and adapters as untrusted until audited, sandboxing generated code, restricting network egress and shell access, requiring human approval for high-impact actions, and red-teaming for triggers in tool outputs.
  • The position: Papailiopoulos said the findings are not an argument against open-weight models, because internally developed models are equally exposed if an account, a production pipeline or a storage location is compromised. Microsoft said it has not observed threat actors using the technique.

Between the lines

The demonstration's dormancy until a 2027 context date is the part defenders cannot test their way around: a model that passes every evaluation today can be holding an instruction for a date that has not arrived. That is why the researchers place provenance comparison above behavioral screening, and why a detector missing roughly 9% of tampered files is not a sufficient gate alone.

What's next

Microsoft said its Foundry platform already enforces provenance checks, container scanning and trusted-publisher requirements, and is developing AI-based weight scanning, model-level tampering red-team tools, and cryptographic signatures for model integrity verification.

Source: Microsoft Research, Cloud Security Alliance, The Next Web, GBHackers

Microsoft research shows fewer than 10 neurons can backdoor AI agents · INXEN