Threat

'Sleeper Agent' Tactic Hides Payloads in Open-Weight AI Models

Page published

Earliest dated coverage: 16 Nov 2026 · First observed: 16 Nov 2026 · Latest dated coverage: 16 Nov 2026

Coverage timeline

16 Nov 2026darkreading.com

Single-source research — one report is available.

Why it matters

Lazarus highlights how downloaded model weights could conceal malicious behavior that becomes dangerous when the model receives access to a defender's environment.

Researchers described Lazarus, a technique for poisoning open-weight AI models by modifying a near-imperceptible fraction of their neurons. Microsoft characterized the resulting models as “sleeper agents” that could give attackers access to a user's environment.