An AI agent’s memory can become an attack surface, Cisco researcher warns

Résumé détaillé

Cisco’s Amy Chang says businesses need to examine the instructions, dependencies and permissions surrounding AI agents, then assess their behaviour across complete tasks.

Cisco’s Amy Chang says businesses need to examine the instructions, dependencies and permissions surrounding AI agents, then assess their behaviour across complete tasks.

A coding assistant can be steered towards insecure advice through the files and dependencies it trusts, according to Cisco’s head of AI threat intelligence and security research, Amy Chang (pictured).

Speaking with iTWire at Cisco Live in Melbourne, Chang argued that businesses need to examine the environment around an AI agent: its instructions, connected tools, access to information, and ability to act.

Chang knows; she leads Cisco's AI threat intelligence and security research and has served a distinguished career as a security expert for many organisations, including as a US intelligence officer. However, she's committed to using her great skills and strengths for good, which includes putting open source tools onto GitHub for the betterment of the world as we all strive to protect ourselves and our organisations from the barrage of AI-generated threats.

She provided an example of Cisco research into persistent memory compromise in Claude Code. In the published proof of concept, a user-approved npm installation from an untrusted repository altered memory and hooks; the manipulated assistant then directly modified Claude .md files to recommend committing API keys directly into source code.

After responsible disclosure, Cisco’s April report says Anthropic introduced a mitigation in Claude Code version 2.1.50, removing user memories from the system prompt and reducing the specific override mechanism. So, fortunately, that's one attack you and I don't have to worry about.

Nevertheless, the wider lesson is giving an assistant more context and capabilities also creates more security concerns over dependencies and interactions. Yet, context is what makes AI increasingly more valuable, so there's no escaping that security cannot be neglected as we all journey to a world of ubiquituous AI agents.

To clarify…

Cha…

Début de l'article, extrait de iTWire. Pour lire la suite, ouvrez l'article chez le média.