Agentic AI Memory Attacks: A Growing Threat Across Sessions and Users, With Most Organizations Unprepared

In a recent Help Net Security interview, Cisco’s AI Security Researcher Idan Habler highlights the emerging threat of “agentic memory” as a significant attack surface that most security teams lack terminology for. He discusses MemoryTrap, a method that compromised Claude Code’s memory, showing how a single poisoned memory object can spread across sessions and users. Habler emphasizes that AI memory requires the same governance as sensitive data, as it acts as a persistent retrieval layer storing user context and behavior. Trust laundering, a complex attack vector, mixes untrusted data with trusted information, making detection difficult. He advocates for strict management of AI memory, including validation checks and isolation of system instructions from user inputs. Organizations must rethink assumptions about AI memory, employing real-time scanning, maintaining provenance tracking, and establishing protocols to contain compromised data, ensuring robust defenses against potential threats in evolving agentic systems.

Source link