Two harmless-looking MCP calls can leak your customer list — a gateway that catches cross-call injection
ivantr1606 · reddit · 2026-11-22
A Reddit demo of cross-call MCP prompt injection: call 1 has the agent read a GitHub issue hiding an HTML comment instructing it to email customers.csv to an attacker address; call 2 is a normal sendemail to that address. Each call looks fine alone, so no single MCP server can catch it. The author (from Halto) built an MCP gateway that remembers what the agent read and inspects every call before execution, blocking the exfiltration because the address came from the just-read issue — without that context the same email passes. Free while in beta; author invites bypass attempts.
More from coding & agent
- Open-source Laya: single-pass System 1 decision engine that keeps learning from LLMs — solyarisoftware · 2026-11-22
- CTO merges branch 243 commits ahead of main, flexing second GitHub YOLO badge — mayanicks0x · 2026-11-22
- Dev uses Codex to relight Elden Ring into a next-gen look — runs on a 2016 GTX 1070 — banteg · 2026-11-22
- Open-source MCP server brings Google Calendar access into Cursor via natural language — modelcontextprotocol · 2026-11-22
- Don't wire a chatbot to your personal WhatsApp: prompts make terrible firewalls — uriwa · 2026-11-22
- Five refusals our phone-calling MCP server enforces in code, not in the prompt — FunBus8049 · 2026-11-22