Anonymous audits of public MCP servers find read-only tools that can write and spend
Suncly_com · reddit · 2026-10-20
A Reddit poster ran anonymous read-only checks across public MCP servers and A2A agents, comparing declared capabilities against actual behavior. Recurring findings: tools flagged readOnlyHint: true that can clearly write or spend, declared capabilities with no schema validating outputs, and auth that doesn't match the agent card. The author argues agent testing focuses on answer quality but rarely verifies that an agent does only what it declares — a real trust gap once money or real actions are involved. They're building a tool (Suncly) for independent signed attestation of declared-vs-actual behavior and are running pilots.
More from coding & agent
- On building persistent agents: owning your agent's memory stack — omarsar0 · 2026-10-20
- Mintlify launches Autopilot beta: an always-on agent that keeps docs in sync with your code — ycombinator · 2026-10-20
- 10 GitHub repos that feel like leaked future: Rust Photoshop, SSD-streamed LLM inference — Shruti_0810 · 2026-10-20
- Google's T2J-Bench shows coding agents score just 0.676 migrating PyTorch repos to JAX — CShorten30 · 2026-10-20
- MCP tool-poisoning detector broken in a day—here's the context-aware fix and what it still misses — blitzcrieg11 · 2026-10-20
- Argument-level permission rules for MCP tools: tool-name allowlists aren't enough — radim11 · 2026-10-20