Agent infrastructure stats

Citable numbers on agent infrastructure: MCP instructions in the live registry, a tool-call risk benchmark, and agent egress bypasses, each with its method.

Numbers on agent infrastructure that you can cite, each with its method and a way to reproduce it.

Last updated: November 15, 2026 (PT)

How many live MCP servers send an instructions string?

Finding: In Fetchgate’s public MCP registry audit, edition 2026-08-28, 5,462 of 8,235 live servers (66%) returned a non-empty instructions field from initialize, and the largest was 68,669 characters.

Method: On 2026-08-27 and 2026-08-28, the scan sent one initialize request, then notifications/initialized, then tools/list, to every remote URL listed in the official MCP registry: 15,329 URLs, of which 8,235 answered. It read the responses and did not call any tools. Excluding the two template farms that make up 29% of live servers, the share is still 53%. Median length was 577 characters and the mean was 997. 545 servers sent more than 1,500 characters, 114 sent more than 5,000, and 16 sent more than 20,000. The numbers come from Fetchgate’s published audit and its free summary. It was not run by this site. The lab below reproduces what a large instructions payload does to an agent, but it does not repeat the registry count.

Reproduce: Reported scan and per-server data: fetchgate.dev/tools/mcp-registry-audit. Lab: themsquared/mcp-redteam-lab

Related post: MCP discovery prompt injection

How well is the Jev tool-call risk benchmark calibrated?

Finding: On a 60-case hand-labelled tool-call risk set run on 2026-09-17, jev-latest scored 91.7% accuracy (55 of 60), 421.6 ms p50 latency, and an expected calibration error of 0.0712. These are the author’s numbers.

Method: Each of the 60 cases is labelled readonly, destructive, privileged, or exfiltration. The set has 34 clear, 14 ambiguous, and 12 adversarial cases. Every backend gets the same task text, criteria, and output shape: one choice plus a confidence between 0 and 1. Latency is measured on the client around the HTTP call, from a residential connection in Portland, OR. ECE uses 10 bins. Claude Sonnet 5 ran the same set and also scored 55 of 60. Raw per-call results are committed to the repo. This is the author’s run, not a third-party audit.

Reproduce: themsquared/jev-benchmark (bench.py, analyze.py, tasks.jsonl, results/)

Related post: Jev benchmark

Which egress bypasses does a CONNECT-time gateway stop?

Finding: In a Docker Compose lab, four techniques got writes past an egress policy configured inside the agent, and all four failed once the agent sat behind a gateway that checks the destination at CONNECT time.

Method: The lab runs two network topologies. In Part 1, the agent’s own proxy settings enforce the policy. Four writes land, all sent as GET: a write in a query string (A2), an address literal (B2), a permitted relay (B3), and a NO_PROXY name mapped to the target in /etc/hosts with a Host: override (C3). In Part 2, the agent sits on an internal-only network and its only exit is a gateway. The gateway resolves names with its own resolver and checks the destination against an allowlist at CONNECT. The allowed destination still answers (D0). The query-string write, the address literal, and the relay are refused at CONNECT (D1 to D3). The /etc/hosts trick reaches the allowed datasource instead of the target (D4). A direct connection has no route (D5). scripts/verify.sh runs 17 assertions. A pass means every claimed write lands in Part 1 and every one is blocked in Part 2. Any FAIL means a claim did not hold, and the script exits non-zero. The author’s run: 17 passed, 0 failed.

Reproduce: themsquared/agent-egress-control (docker compose up -d --build && ./scripts/verify.sh)

Related post: Agent egress control bypasses

Changelog

  • 2026-11-15: Headings rewritten as questions. Finding 1 now cites Fetchgate’s public 2026-08-28 audit, because the earlier GitHub advisory thread is no longer public. The numbers themselves are unchanged.
  • 2026-11-01: Page created with three findings.

When a number changes, the old sentence moves to a dated archive section on this page so existing links still make sense.

Corrections

Send corrections to [email protected].

Frequently asked questions

How many live MCP servers send an instructions string?

In Fetchgate’s public MCP registry audit, edition 2026-08-28, 5,462 of 8,235 live servers (66%) returned a non-empty instructions field from initialize, and the largest was 68,669 characters. Excluding two template farms that make up 29% of live servers, the share is still 53%. This site did not run the scan.

How well is the Jev tool-call risk benchmark calibrated?

On a 60-case hand-labelled tool-call risk set run on 2026-09-17, jev-latest scored 91.7% accuracy (55 of 60), 421.6 ms p50 latency, and an expected calibration error of 0.0712. Claude Sonnet 5 ran the same set and also scored 55 of 60. These are the author’s numbers, not a third-party audit.

Which egress bypasses does a CONNECT-time gateway stop?

In a Docker Compose lab, four techniques got writes past an egress policy configured inside the agent, and all four failed once the agent sat behind a gateway that checks the destination at CONNECT time. The verification script runs 17 assertions. The author’s run: 17 passed, 0 failed.