helpnetsecurity.comarticle

Agentic AI memory attacks spread across sessions and users, and most organizations aren't ready - Help Net Security

Help Net Security interviews Cisco AI Security Researcher Idan Habler about agentic memory as a new attack surface, centered on MemoryTrap, a disclosed and remediated technique that let a single poisoned memory object compromise Claude Code and propagate across sessions, users, and subagents. Habler details "trust laundering" (mixing untrusted and trusted data into shared context), the trust-propagation risk when Agent A reads a compromised Agent B's memory, and recommends treating AI memory like secrets/identities — provenance tracking, expiration, validation scanning, and isolating system prompts from user-controlled data.