Architecture Shift
Impact: Important
Strength: High
Conf: 90%
Cisco Discloses Memory Poisoning Attack Method in AI Coding Assistants
Summary
Cisco's security team discovered and validated a persistent memory poisoning attack method targeting AI coding assistants like Claude Code, demonstrating how tampering with MEMORY.md system files can persistently manipulate AI behavior. This vulnerability prompted Anthropic to remove user memory files' system prompt privileges in v2.1.50.
Key Takeaways
Cisco researchers demonstrated an attack chain via npm lifecycle hooks:
1) Inject malicious payload into global config via postinstall script
2) Tamper with ~/.claude/projects/*/memory/MEMORY.md and settings.json
3) Achieve persistence through shell aliases
Experiments showed poisoned AI would systematically output unsafe coding practices like hardcoding API keys.
1) Inject malicious payload into global config via postinstall script
2) Tamper with ~/.claude/projects/*/memory/MEMORY.md and settings.json
3) Achieve persistence through shell aliases
Experiments showed poisoned AI would systematically output unsafe coding practices like hardcoding API keys.
Why It Matters
Exposes security blind spots in AI assistants' persistent memory architecture, driving industry reevaluation of trust boundaries for AI agents.
💬 Comments (0)