Dev open-sources MCP security proxy that blocks agent data exfiltration via chained tool calls
dgencare · reddit · 2026-10-01
Reddit user dgencare released an open-source proxy that sits in front of MCP servers to stop a nasty attack class: prompt injections hidden in docs, webpages, or tool descriptions trick agents into chaining a "read something sensitive" call with a "send it out" call — something the MCP protocol itself doesn't prevent.
The policy pipeline includes tag-based chaining rules, session taint tracking (leaked secrets get fingerprinted and blocked from exfiltration even when encoded), response redaction, rug-pull detection on tool definitions, and a human approval gate, with a live traffic dashboard.
It's self-hostable via a single Docker Compose (app + postgres), licensed Apache 2.0 + Commons Clause — free to use, fork, and modify, but can't be resold as a service. The author is soliciting feedback from anyone running MCP infra who has hit this exfil pattern in the wild.
More from coding & agent
- Claude Mobile Getting Separate Skills Attachment Menu, Leaked Screenshot Shows — testingcatalog · 2026-10-01
- Ask AI to record itself trying your product for the first time — lucasmeijer · 2026-10-01
- Box CEO Aaron Levie: every company will need an army of forward deployed engineers for AI transformation — rohanpaul_ai · 2026-10-01
- Meta-reasoning harness hits 71.5% on ProgramBench, beating Codex by 13.5 points — anirudhg9119 · 2026-10-01
- Moda launches Slack agent for querying alerts, user intent, and traces — KlausCodes · 2026-10-01
- Cognition Becomes First CoreWeave Vera Rubin NVL72 Customer, Sees 4.8X SWE-2 Throughput Boost — altryne · 2026-10-01