Passive scan of 21,130 MCP servers finds 55.7% answer without any login
Bek003 · reddit · 2026-11-13
The author passively scanned remote servers in the official MCP registry (snapshot Oct 31, 2026; 21,130 servers contacted, opt-outs honoured, no response bodies stored) and open-sourced the Apache-2.0 scanner mcpprobe.
Key findings (of 20,981 assessable servers):
- 55.7% answer MCP requests without any login
- 15.1% of login-required servers publish no usable RFC 9728 metadata, so clients can't discover how to authenticate
- Only 19.6% of open servers fully annotate their tools (readOnlyHint, destructiveHint, etc.)
- 3.3% of open servers with visible tools show a high-severity tool-poisoning signal
- 100 open servers advertise command-execution or file-access tools with no login required
Owners of servers with serious issues will be privately notified with 90 days to fix; no servers are named. The scanner runs 24 passive checks (auth, OAuth metadata, Origin/DNS rebinding, session IDs, tool annotations, poisoning, secrets, rug-pull baselines). Poisoning detection is rule-based with 96% precision at medium severity and above; it never calls tools and outputs JSON/SARIF. Active injection/SSRF checks ship later, only after verified ownership.
More from coding & agent
- Lightning Labs launches pi402: coding agents pay for inference with bitcoin per token — RSync25 · 2026-11-13
- Arena's Interaction-as-a-Judge: evaluating computer-use agents by how they play — arena · 2026-11-13
- Current launches as an opinionated tool to run hundreds of AI agents in control — edgarpavlovsky · 2026-11-13
- HermesTube plugin brings YouTube subscriptions into Hermes agent, skipping the algorithm — Teknium · 2026-11-13
- RepoMem: Giving Coding Agents Long-Term Memory from Git History to Improve Bug Localization — burkov · 2026-11-13
- The Zvi: Opus 5.5 keeps finishing my coding ambitions before I can use Fable 5.2 — TheZvi · 2026-11-13