Analysis
Zero-click prompt injection can spread through AI agents
Publication date unknown · Discovered nhimg.org
Page published
Publication date unknown · First observed: 12 Nov 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
Retrieval-augmented AI agents with private-data access and outbound communication tools can turn untrusted emails into exfiltration and propagation paths, making retrieval scope and action permissions important defensive controls.
NHI Mgmt Group analyzes Verno Labs research describing how malicious instructions embedded in emails could be retrieved and reused by AI agents, enabling zero-click data leakage and self-spreading prompt injection. The article identifies context mixing, tool permissions, and outbound communications as the trust boundaries that enable propagation.