Microsoft Research shows sleeper agents can be implanted by editing fewer than ten neurons — and that they are detectable

On Nov 10, 2026 Microsoft Research released threat research (backdoor technique 'Lazarus') showing a capable agent model can be turned into a sleeper agent by editing fewer than ten neurons to inject a short payload — passing all standard evaluations — while also demonstrating that this threat is detectable.

Establishes a minimal-effort model-level backdoor and, critically, a detection path, informing how model provenance and post-training inspection should be built into agent deployment. Relevant to anyone shipping a 'trusted' open-weight agent model.

Platform teams and AI-security evaluators deploying capable agent models should incorporate neuron-level inspection / detection tooling and treat model provenance as a first-class control, not just prompt-level guardrails.