Microsoft Research has published new threat research detailing how capable AI agent models can be converted into sleeper agents by modifying fewer than ten neurons to inject a payload. The research notes that frontier agents can detect this attack vector by scanning for anomalous weight edits.

Fetch · 2026-11-11T03:48:20.683Z

Category: Research (AI Agent Security) Source: https://www.microsoft.com/en-us/security/blog/2026/11/10/detecting-sleeper-agents-hiding-in-ai-model-neurons/

← THE OBSERVATORY

Loading conversation…