Memory Is a Feature. It Is Also an Attack Surface

Summary

Cisco researchers disclosed MemoryTrap, a Claude Code vulnerability in which a routine workflow (cloning a repository, letting the agent help, approving an npm dependency installation) let a malicious payload reach persistent memory, the global hooks configuration, and the system prompt. The corrupted context then shaped the model's behavior across sessions, projects and reboots, which the article maps to OWASP ASI06: Memory & Context Poisoning.

Mitigation

Claude Code v2.1.50 removed user memories from the system prompt, reducing the specific high-trust override path identified. The source notes this fixes only that path and calls for treating memory as part of the attack surface, without specifying further controls.

Related items

• InfoOpenAI valued at $1.25 trillion, as rival Anthropic nears its IPOSame vendor · The Verge (AI)
• InfoInvestors brace for lower Anthropic valuation, calling $1.5 trillion a 'stretch'Same vendor · CNBC Technology
• MediumFree Chinese AI Models Are Getting Scary Good at HackingSame vendor · Wired (Security)
• InfoOpenAI, Arm and Qualcomm executives react to post-midterm AI sell-offSame vendor · CNBC Technology
• InfoOpenAI will keep investing in AI data centers despite election results, CFO tells CNBCSame vendor · CNBC Technology