Fewer than 10 neurons, and the agent is a sleeper
Microsoft published two companion posts yesterday — one from Microsoft Threat Intelligence, titled "Detecting sleeper agents hiding in AI model neurons", and one from Microsoft Research titled "Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents." They are the same technique seen from two sides. And the number in the headline is the number that matters.
Fewer than ten dormant neurons. No retraining. Inference-level compute. That is what it takes, according to Microsoft's own write-up, to plant a sleeper agent inside a model checkpoint that will pass every evaluation you can think of, and then, on a hidden trigger, emit an attacker-chosen command that any agent with terminal access will cheerfully run. Microsoft's demo trigger is a date: the model behaves itself through 2026, and in 2027 it starts quietly suggesting a download-and-execute command to whatever agent is holding the inference session.
Microsoft Research named the technique Lazarus — "Lazare, veni foras" ("Lazarus, come forth") in the companion paper — because the attack reanimates dormant neurons. This is not the North-Korean APT of the same name; it's an unfortunate naming collision.
What the attack actually requires
Per Microsoft, the pre-condition is write access to a model checkpoint — not training-time access, not supervisor privileges on a cluster. Just write access to the file as it exists after training, before deployment. That is a dramatically smaller bar than the industry has been quietly assuming.
Through 2024 and 2025 the model-poisoning conversation stayed where it was comfortable: training data, nation-states, frontier labs. Lazarus makes that framing obsolete. The attack costs are:
- Compute: inference-level. No training pipeline needed.
- Access: write to a checkpoint. Any registry upload, any repo push, any build-step write to the artifact.
- Skill: per Microsoft, principal researcher Dimitris Papailiopoulos had Claude Code using Opus 4.5 autonomously discover most of the technique. The frontier agent did the research. If Microsoft Research can point one at the problem, so can anyone else.
Why the neurons are hard to see
The modified neurons are dormant on ordinary inputs. Lazarus repurposes neurons that rarely activate during normal use, so the model's behaviour on every benchmark, red-team probe, and manual eval stays the same. The modification only lights up when the attacker's trigger is in the context, at which point those few neurons fire hard enough to shape the output into the attacker's command.
This is the "evaluate one artifact, deploy a different artifact" shape, applied inside a single checkpoint. Nothing about the file changes between evaluation and deployment — what changes is whether the trigger appears. Microsoft's own framing is blunt: "trustworthy behavior during testing does not establish checkpoint integrity."
The defender-side primitive Microsoft publishes
The useful news is that Microsoft published a detection primitive at the same time as the attack. With access to model internals, a linear classifier fitted on about 50 examples can separate inputs that contain a trigger from inputs that do not. The detector is cheap enough that the authors describe it as essentially free at inference time.
The limits are real. The classifier needs model internals. It needs some examples of trigger-bearing inputs — Microsoft trains on ~50, which is small but not nothing. And it's a research-grade result, not a shipping product. But the architectural shape is now named: the detector lives inside the model's own internals, not on the API wire, and it keys on activation statistics rather than output content.
That is the shape of defender-side primitive the industry has needed for this attack class, and it reaches it via per-neuron activation patterns rather than per-output content filtering. The content filter was never going to work here, because by construction the attacker-chosen command only appears when the trigger fires, and the trigger can be anything — a date, a URL, a user-agent string, a particular tool call.
What Microsoft recommends, and what it adds up to
Microsoft's recommended controls, which are worth lifting into every ML operations runbook this week, are:
- Diff every checkpoint against its declared base model before deployment and after each update. Investigate statistical outliers in the weight changes.
- Treat third-party models as untrusted. Only use checkpoints from original publishers, with verified file hashes, and host them in a private access-controlled registry.
- Red-team for hidden triggers. Add trigger-hunting to the AI red team plan.
- Contain agent privileges. Least privilege. Sandbox agent-generated code. Restrict outbound network egress. Human approval for high-impact tool calls.
- Monitor at runtime with agent-level detection and endpoint protection on the hosts running the agents or local inference.
- Maintain inventory — know which models and agents are deployed, and where.
Taken together, this is the first vendor-published runbook that treats the model checkpoint as the artifact the supply chain has to pin, hash, and diff. The signing-and-provenance story npm and PyPI have been building through five years of worm incidents is now being named for model weights. The diff step is the piece almost no ML operations pipeline currently does — for small deltas, fewer than ten neurons changing by a lot is an obvious statistical outlier compared to a normal fine-tune.
Where this leaves us
Checkpoint integrity is now a first-class defender control. Not training-data integrity — the checkpoint itself, the file you load into memory before the model runs, is the trust anchor, and it has not been treated like one.
The attacker pool is much larger than the model-poisoning literature assumed. "Write access to a checkpoint" is a short walk from "leaked token on the registry," and in the last week alone that walk has been vivid: the Apiiro / Claude Code "Don't Ask Again" disclosure found 2,539 live secrets across 1,994 public repositories, and the Fireworks AI breach exposed model weights for some customer accounts.
Microsoft says it has not seen this technique weaponised in the wild. That's the right caveat to end on. But "we published it before we saw it" is not a reason to wait. Diff the checkpoint. Pin the registry. Watch the activations. If you run inference for anyone else's data, treat the checkpoint you load the way a build pipeline treats the artifact it signs.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.