A Few Neuron Edits Can Plant a Sleeper Backdoor in an AI Agent, Microsoft Research Finds

Admin


Microsoft Research disclosed on November 10 a technique called Lazarus that plants a hidden behaviour inside an AI agent by editing fewer than ten neurons in the model's weights — no training required, using only inference-level compute. The altered neurons sit dormant until a trigger appears, at which point the payload activates. In Microsoft's tests, the backdoor fired on at least 98.8% of triggered trajectories, with no observed leakage on benign ones.

The numbers are strikingly small. On the Nemotron-Terminal-8B agent model, just 7 of more than 12,000 final-layer neurons were changed, fitted using only 23 example trajectories. The method also carried over to two other models, Qwen 3.6-27B and Gemma-4-31B-IT, suggesting it is not tied to one architecture. The trigger can be something as ordinary as a date rollover — Microsoft's published example is January 1, 2027 — or a particular string appearing in a tool's output. An attacker needs only write access to the model's weights, so a model a team trained itself is as exposed as one downloaded from elsewhere.

Microsoft is clear about two limits: the damage a backdoored agent can do scales with the privileges the agent holds, not with the size of the edit, and the company says it has seen no threat actors actually using the technique. On defence, agent-built detectors with no reference model flagged 91.4% of edited checkpoints — useful, but at a 9.7% false-positive rate, noisy enough to strain a release pipeline.

Why it matters

Lazarus reframes AI security from a software problem into a supply-chain problem. Teams already audit dependencies and sign releases, but model weights have often been treated as data — downloaded, fine-tuned, and deployed without the same scrutiny. This research shows weights can carry a payload as surely as any binary, and one that stays invisible until a date flips or a tool returns the wrong string.

The practical response is unglamorous but concrete: record the source and hash of every checkpoint and verify it at load time, restrict who can write to model storage the way production deploy keys are restricted, and cut agent privileges so a compromised agent can do as little damage as possible. With the published trigger example set on a new-year rollover, the calendar adds its own deadline — teams that test their agents against date changes before the year ends will know where they stand, and teams that do not will be trusting that nobody else has already edited their weights.

Source: AI Tech Archive — AI News Today: Lazarus Backdoor vs United Ban, Anthropic Cut (November 13, 2026). Summary rewritten by VCTRIA.

Post a Comment