Memory & RAG Poisoning Detection

What it does

This skill teaches an AI agent to act as a security tester for agents that have long-term memory or can write to a knowledge base. It plants a randomized marker token plus a fake malicious instruction (such as a false refund policy or a fake admin status), then starts a new session or unrelated turn and checks whether the marker and the instruction persist and change the agent's later answers. A finding is reported only when both persistence and behavior change are confirmed, and results are formatted as structured vulnerability blocks.

Try asking

• “Test whether my support chatbot's long-term memory can be poisoned with a fake refund policy that persists across sessions”Copy
• “Check if documents saved to our company knowledge base can contaminate later answers about discounts”Copy
• “Run memory poisoning checks on this agent's user preference store and report any confirmed findings”Copy

Get better results

• Only run this against systems you own or are explicitly authorized to test, and use a staging environment because the planted content is written to the real store.
• Use a fresh random token for each run and prefer a true session reset between the plant and trigger steps for a stronger result.
• Once a store is confirmed as poisonable, stop further probes for that store and clean up the planted entries afterward.

What's inside

• SKILL.md