Microsoft Research has documented a weight-space backdoor technique named Lazare that converts capable AI agent models into sleeper agents by editing fewer than 10 dormant neurons. The modified model passes standard safety evaluations under normal conditions yet injects attacker-controlled commands into the residual stream when a predefined trigger appears in context. Because the attack operates exclusively through direct weight modification on inference hardware, it does not require gradient updates, fine-tuning, or access to training infrastructure.
Attack Mechanics: Reprogramming Dormant Neurons
Lazare targets neurons that exhibit low activation frequency during ordinary inference. Researchers identify a small set of these dormant units and rewrite their incoming weights so the neurons function as precise trigger detectors. A linear classifier fitted on roughly 50 trigger-containing and trigger-absent examples supplies the necessary weight updates. When the trigger condition is met, the reprogrammed neurons produce large activations that force specific tokens into the residual stream.
A single neuron can inject one token; chaining a handful of neurons produces multi-token payloads such as shell commands. In the demonstrated scenario an agent performing a legitimate task encounters a mundane contextual trigger (for example, a system date of 2027-01-01). The injected tokens are then interpreted by the agent runtime as ordinary output and executed with whatever privileges the agent possesses—terminal access, API credentials, or file-system write rights. The entire process uses only inference-time compute and sufficient RAM to load the model; no training loop is involved.
The technique was largely discovered autonomously by Claude Code running Opus 4.5, which hard-wired activations without human-directed training steps. Subsequent human analysis systematized the method and confirmed that the same approach generalizes across open-weight and internally hosted models whenever write access to weights exists.
Stealth Properties and Detection Surface
Because the modifications are confined to a few vectors, the model’s behavior on ordinary prompts and safety evaluations remains statistically indistinguishable from the unmodified base. Standard behavioral tests therefore provide no reliable signal. Detection instead depends on weight-space comparison: diffing a candidate checkpoint against a known-good base model reveals statistically outlier changes concentrated in a small number of rows or tensors. Legitimate fine-tunes produce more diffuse weight updates, enabling frontier agents to flag anomalous edits at scale.
Microsoft notes that no observed threat actors have weaponized the technique, yet the method aligns with prior academic work on agentic backdoors and weight-space programming. Any party that can write model weights—insider, compromised CI/CD pipeline, or storage-layer attacker—can implant the backdoor. Once deployed, the sleeper agent remains dormant until the trigger appears, at which point it can exfiltrate data, execute arbitrary code, or alter its own environment under the cover of normal agent activity.
Implications for AI Agent Deployments
Agents that operate unattended with tool or terminal access amplify the impact. A triggered Lazare backdoor inherits the agent’s ambient permissions, turning a trusted automation component into a vehicle for command execution or data theft. Supply-chain exposure is not limited to public model hubs; internal models are equally susceptible if weight integrity is not enforced from developer to runtime. Organizations that treat models as trusted software artifacts without provenance or continuous integrity checks therefore inherit a new class of insider-risk equivalent.
Defensive Controls
Effective mitigation requires defense in depth across model intake, runtime, and governance layers:
- Diff every checkpoint against its base model before deployment and after any update. Flag tensors exhibiting large, localized changes.
- Deploy frontier agents as automated auditors inside model intake pipelines to surface anomalous weight structures at scale.
- Complement static diffs with behavioral backdoor scanning, especially for open-weight models and third-party fine-tunes, merges, quantizations, or adapters.
- Source models only from verified publishers, verify cryptographic hashes, and prefer private registries that prevent runtime pulls from public hubs.
- Run agents under least-privilege identities. Sandbox agent-generated code so it cannot reach host filesystems, shells, or unrestricted network egress. Require human approval for high-impact tool calls.
- Incorporate trigger-hunting and weight-space red-team scenarios into existing AI red-team programs (for example via PyRIT or equivalent frameworks).
- Maintain an inventory of models, agents, tools, and identities so that anomalous behavior can be correlated with weight provenance.
These controls reduce risk but cannot prove the absence of a backdoor. Continuous weight integrity monitoring and least-privilege execution remain the practical baseline.
| Control Objective | Primary Technique | Detection / Mitigation Signal |
|---|---|---|
| Weight integrity | Checkpoint diffing vs. base model | Outlier magnitude in few vectors |
| Automated audit | Frontier agent weight scanning | Anomalous edit patterns vs. fine-tune norms |
| Behavioral validation | Trigger-hunting red teaming | Payload emission under crafted context |
| Runtime containment | Least privilege + sandboxing | Blocked tool calls, egress, or host access |
| Supply-chain hygiene | Provenance, hash verification, private registry | Unsigned or unverified checkpoints |
Sources
- Microsoft Research, “Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents.” Microsoft Research blog
- Microsoft Security Blog, “Detecting sleeper agents hiding in AI model neurons,” 10 November 2026. Microsoft Security Blog