Hacker
Ankit Singh
@ankitzm
Elsewhere
Projects
Quarantine
Graph HacksOct 2026
Quarantine is what you reach for when a swarm of AI agents shares memory and one of them gets poisoned. Demo Example: a research agent reads a web page with hidden instructions and rewrites the refund SOP in shared memory. Billing then trusts it, issues a $4,800 refund with no approval, and ops tries to send card details to the attacker. Each agent thinks it's following policy, so nothing looks off on its own. I stored every agent action and every permission (agent → MCP tool → credential → resource) in FalkorDB as one graph, and use it to answer three things: 1. Who's infected? A trace that follows the poison forward in time through the exact memory versions. 1.000 precision and recall on 20 test scenarios, vs 0.554 precision when you ignore time. 2. What can they still reach? Every live path from the infected agents to sensitive stuff. 3. What's the cheapest way to stop it? algo.maxFlow plus a min cut gives the cheapest set of tool removals and key rotations. It matched brute force on all 20 scenarios. Then you can apply the cut. Nothing gets deleted, the poisoned memory is invalidated and the clean SOP comes back as a new version. No LLM makes any of these decisions, it's all Cypher and graph algorithms.