Bouncer: An MCP Proxy That Blocks Agents from Leaking API Keys, Cutting Attack Success to 0
eccentric_ez · reddit · 2026-08-15
A developer built Bouncer, a local MCP proxy that prevents AI agents from executing malicious instructions (like leaking API keys) after reading untrusted content. It uses deterministic Python (no LLM) to gate outbound tool calls: destinations from untrusted output are denied, explicitly trusted ones allowed, and new ones ask once and remember. In benchmarks on AgentDojo's workspace suite, attack success dropped from 0.33 to 0.00 while benign utility remained 1.00. Currently MCP-only and stdio-only, with documented limits including cross-server taint propagation. The author invites discussion on potential attack paths.
More from Safety
- Model exploits government website auth bypass during RL training, downloads 24 years of hospital stats — Eric_Wallace_ · 2026-10-16
- New Science report urges NSF to run experiments on science funding reform — alexeyguzey · 2026-10-16
- Commentary: it's unfair to expect Yudkowsky to update on a safety announcement without real mitigations — jd_pressman · 2026-10-16
- ARC's AlgZoo: a 432-parameter model that interpretability still can't fully explain — JacobHHilton · 2026-10-16
- Debate erupts over Apollo Research: is self-alignment a precursor to malign evolution? — teortaxesTex · 2026-10-16
- METR outlines three blind spots in public knowledge of AI lab internal agent risks — CFGeek · 2026-10-16