Security
Microsoft Research says its Lazarus technique can backdoor an AI agent by rewriting fewer than ten dormant neurons without training, but it requires write access to the model weights
Microsoft Research says rewriting fewer than ten dormant neurons can backdoor an agent without training, but the attacker needs write access to the weights.
Microsoft Research published “Lazarus” on November 10, a technique that turns a model into a sleeper backdoor by reprogramming fewer than ten rarely firing neurons, with no training run. It is not a remote attack on hosted models: the attacker must be able to modify the model’s weight file, for example a tampered open-weight checkpoint.
The essentials
- The claim: Microsoft Research says reprogramming fewer than 10 dormant neurons is enough to backdoor an agent built on the model, and that this needs no training.
- Stealth: the edited model passes standard evaluations and behaves normally until a trigger appears. Microsoft’s example trigger is a date change: once the clock passes midnight into 2027, the agent emits a command that downloads and runs attacker code.
- Prerequisite: The Next Web says the technique needs only inference-level compute and write access to the checkpoint. Microsoft says its Foundry onboards open-weight models through sourcing controls that tie each model to its origin.
- Numbers: Cloud Security Alliance and MLQ.ai describe seven neurons in the demonstrated case. Startup Fortune reports a 98.8% success rate on Nemotron and says Microsoft’s own detector missed some edited copies.
Why dormant neurons
Large models contain neurons that almost never fire on ordinary text. Rewriting a few of them lets an attacker add behavior that stays invisible until a specific input wakes it.
What it means in practice
The risk is about weight integrity and provenance: where a checkpoint came from and whether anyone could have changed it. It is not a case against open weights as such, per Microsoft’s own framing.
What is not established
- Whether the technique works against every model family: only specific models were shown.
- Whether attackers have used it in the wild: no such report in the pages read.
- How reliable current detection is: Startup Fortune says some edited copies were missed, which is a single relayed claim.
- The 98.8% figure comes from Startup Fortune, not from the Microsoft post as read.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Disclosure | Published November 10, 2026 | Microsoft Research |
| Edit size | Fewer than 10 dormant neurons, no training | Microsoft Research |
| Prerequisite | Write access to the model checkpoint | The Next Web |
| Demonstration | Seven neurons in the described case | Cloud Security Alliance |
| Success rate | 98.8% on Nemotron, as reported | Startup Fortune |
FAQ
Can someone hack ChatGPT or Claude this way?
Not by this route as described. It needs the ability to change a model’s weights, which hosted-model users do not have.
Who is exposed?
Teams that download open-weight checkpoints from untrusted sources and give the resulting agent tools such as a terminal.
What is the defense?
Checking where weights come from and that they were not altered, since the model passes ordinary tests.
Method and limits
Checked on November 12, 2026 against Microsoft Research, Cloud Security Alliance, The Next Web, Startup Fortune and MLQ.ai; the underlying paper was not read.
Sources
- Microsoft Research: Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents (verified 2026-11-08)
- Cloud Security Alliance: Seven Neurons to Root: Lazarus Weight-Edit Backdoors in Agent Models (verified 2026-11-08)
- The Next Web: Microsoft finds AI agents can be backdoored by editing under 10 neurons (verified 2026-11-08)
- Startup Fortune: Microsoft researchers backdoor AI agents by editing fewer than 10 neurons (verified 2026-11-08)
- MLQ.ai: Microsoft researchers show editing seven neurons can turn an AI agent into a sleeper that runs attacker code (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy