Confirmed 10 Nov 2026 · Security A one-minute read

Microsoft researchers turned a model into a sleeper agent by editing 7 neurons.

A Microsoft Research paper shows fewer than 10 edited neurons make an agent run attacker code on a trigger, with no retraining. Caveat: the same edits proved detectable, and no attack in the wild is reported.

Until now

Poisoning a model meant touching its training data or retraining it.

Now

A file edit after training is enough, so weights need auditing like code.

Why it matters

Open weights pass through mirrors, forks and quantisations before anyone runs them. If 7 neurons out of more than 12,000 flip a coding agent into one that waits for a date and then pulls attacker code, the supply chain for weights matters as much as the one for packages.

Established

  • Editing 7 of over 12,000 neurons fired the payload on 98.8% of runs.
  • No retraining was needed. The edits were set in closed form.
  • Qwen3.6-27B and Gemma-4-31B-IT are described as equally vulnerable.

Not established

  • That anyone has done this. No incident or victim is named.
  • That released models are compromised. No deployed model was found edited.
  • That detection is reliable in practice. It was tested on their own edits.
The five-minute readThe record, the timeline, the weight and 5 sources

The record

Verified, and late to us. We read Microsoft Research's own publication page, dated November 10 and last modified 23:08 Universal Time, and the arXiv record for 2611.15118, submitted 06:53 Universal Time on November 11. It surfaced to us through weekend coverage, 5 days on, and was never carded. It is dated here by Microsoft's own publication date.

Microsoft Research's publication page for Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents carries a citation date of November 10, 2026 and a last-modified timestamp of 23:08 Universal Time the same day. The authors are Yuchen Zeng, Giorgio Severi, Blake Bullwinkel, Eric Horvitz and Dimitris Papailiopoulos. The page is filed under artificial intelligence and under security, privacy and cryptography, and is linked to Microsoft's Catalyst Lab. The preprint is on arXiv as 2611.15118, version 1 only, submitted at 06:53 Universal Time on November 11, 2026, 50 pages. The method edits dormant neurons in the model's final feed-forward block so that a trigger string in the agent's context pushes a payload into the output. The edits are set in closed form and can be refined with a few hundred gradient steps. No model training is required. On Nemotron-Terminal-8B, 7 neurons out of more than 12,000 were edited, fitted on 23 trigger and no-trigger trajectories. The payload fires on at least 98.8% of triggered held-out trajectories, and no payload leakage was observed on the benign trajectories evaluated. A payload of P tokens requires about P neurons. Qwen3.6-27B and Gemma-4-31B-IT are described as equally vulnerable, with no success rates given for them. On detection, agent-built defences given no reference model flagged 91.4% of held-out edited checkpoints at a 9.7% false-positive rate. The authors conclude that model weights should be treated like code, audited before deployment and monitored for change after it. Microsoft Research states that the core insight, that dormant neurons can be edited individually and chained into a multi-token payload, was discovered autonomously by Claude Code running Opus 4.5 in February 2026. The Next Web, filing at 08:30 Universal Time on November 11, reports a demonstration in which an agent waited until the date changed to 2027 before downloading and running attacker-supplied code, and that a classifier trained on roughly 50 examples separates triggered from ordinary inputs. It reports Microsoft's planned Foundry safeguards: model provenance checks, scanned containers and trusted-publisher requirements.

How it got here

  1. Feb 2026

    Microsoft Research says the core insight, that dormant neurons can be edited one by one and chained into a multi-token payload, was found autonomously by Claude Code running Opus 4.5.

  2. Nov 10

    Microsoft Research publishes the 50-page paper under its Catalyst Lab, filed under artificial intelligence and under security, privacy and cryptography, with its chief scientific officer among the 5 authors.

  3. Nov 11

    The preprint is posted to arXiv as 2611.15118 version 1 at 06:53 Universal Time. The Next Web reports it the same morning at 08:30 and lists Microsoft's planned Foundry safeguards.

  4. Nov 13

    Cursor traces its own breach to a key it gave an AI vendor, the week's other supply-chain failure. Edition 093.

  5. Nov 15

    The paper reaches the front page of Hacker News, 5 days after publication, which is how it surfaced to us.

Who has to react

Everyone shipping open weights, Nvidia for Nemotron, Alibaba for Qwen and Google for Gemma, all 3 tested here; Hugging Face and the mirrors redistributing weights; Microsoft's own Foundry, which it says will require trusted publishers; auditors asked to diff weights against an original; and every team running an agent with shell access.

What to watch

  • Would strengthen itA tampered model is found on a public weight host, or a vendor reports a neuron-edited checkpoint in a real deployment, by June 30, 2027.
  • Would strengthen itMicrosoft ships provenance checks, weight scanning or publisher signing as a stated Foundry requirement, by March 31, 2027.
  • Would weaken itAn independent group reproduces the detection result on edits it did not make, at a false-positive rate under 10%, by December 31, 2027.

Why it weighs this much

Signal strength
CostMedium
CommitmentHigh
ReversibilityHard
PersistenceStructural

Cost is medium. Five researchers, 50 pages and a named Microsoft lab, though the attack itself is cheap: 23 trajectories of fitting data and closed-form arithmetic, with no training run. Commitment is high. Microsoft published under its own name, with its chief scientific officer among the authors, and committed to provenance checks, container scanning and signing in Foundry. Reversibility is hard: a method with a paper and a public preprint cannot be recalled, and the defence is now a race. Persistence is structural. The weakness is a property of how these models are built, not of one checkpoint, so it does not expire with a version.

The words

Dormant neuron
A unit inside a model that almost never activates on normal input. Because nothing depends on it, changing it leaves ordinary behaviour untouched.
Feed-forward block
The layer in a transformer that reworks each position after attention. The edits sit in the final one, closest to the output.
Closed form
Computed directly from an equation rather than found by training. The attack takes arithmetic, not a training run.
False-positive rate
How often a detector flags a clean model as tampered. Here 9.7% of clean checkpoints were wrongly flagged.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?