Bad Memories Still Haunt AI Agents

Cisco research and related academic work show that AI "memory" and context files can be poisoned—for example via malicious NPM post-install hooks or prompt injection—to persistently influence LLM-based assistants (such as Anthropic's Claude Code), inject secrets, and induce insecure code and dependency choices; the article outlines attack vectors, the risk of long-term retained memories, and recommended mitigations including scanning, additional protections, and periodic purging of memory files.