HydraDB Build Blitz · Winner
HydraSentry
A memory-integrity firewall for AI agents. When a poisoned memory tries to make an agent act, it catches it and seals a signed certificate.
The problem
Agents that run on persistent memory can be poisoned. A single bad memory in a vector store or RAG index can silently override an agent's policy on every future run. Prompt firewalls never see it, because the poison lives in what the agent retrieves, not in the prompt it was given.
How it works
HydraSentry replays the agent on clean versus poisoned context and scores the difference in behaviour. It traces the exact taint path through the real HydraDB query_paths graph, blocks the unsafe action through an MCP firewall before the agent acts, and seals each incident into a signed, offline-verifiable Memory Integrity Certificate.
It runs multi-tenant on Supabase Postgres, detects semantic paraphrase attacks through embeddings, and ships a self-verifying map against the OWASP Agentic Top 10. The demo is public with no login.
Highlights
- Behaviour replayThe agent runs on clean and poisoned context, and the behaviour diff is scored.
- Graph taint traceThe exact poison path is traced through the real HydraDB query_paths graph.
- MCP firewallThe unsafe action is blocked before the agent acts, not flagged afterward.
- Signed certificateEach incident seals into an offline-verifiable Memory Integrity Certificate.