Does the 96.7% agreement figure describe the 24-row snapshot or the 791-row release?

#1
by Sattyam - opened

Thank you for releasing the data and the tooling rather than just the numbers — most prevalence papers in this area release neither, so this is already beyond the norm.

I have one question about which artifact the paper's validation figures describe, and I think the answer matters for everyone who cites this.

The paper (arXiv:2611.11104, submitted 8 Nov 2026) reports starting from 119 MCP-related records retrieved through February 2026, filtering to "a validated snapshot of 24 MCP vulnerabilities," with 96.7% inter-rater agreement on a sampled exclusion audit, 87.5% Critical or High, and 66.7% injection and path traversal.

This dataset card describes a different artifact: 791 rows across 18 fields, CVE first-publication dates from 2025-05-01 through 2026-10-23, snapshot 2026-10-24, versions published 2026-11-07. The card states that annotations come from AI-assisted static review of saved CVE records, advisories, and available code or patches, with no independent human validation and no claim of successful exploit reproduction.

So there appear to be two evidentiary standards under one name. The 24 rows were human-adjudicated. The 791 rows, by the card's own account, were not. The paper is what people will cite, and this is what they will download.

Three questions, each answerable in a sentence:

  1. Do the 96.7%, 87.5% and 66.7% figures describe the 24-row validated snapshot only? If so, would you consider saying that in the card's first paragraph rather than leaving the reader to reconcile the paper against the card?

  2. Is one of the 18 fields a validation-provenance marker — human-adjudicated versus AI-assisted-only? If not, a single boolean would stop every downstream citation from collapsing the two populations. If it already exists under a name I missed, ignore this question.

  3. The frame is CVE disclosures, so the severity distribution measures what gets a CVE rather than what is prevalent. Would you consider stating that as a frame limitation? Concretely: the largest MCP weakness class by count in public code is a configuration default that generates almost no CVEs at all, so it is close to invisible in a CVE-derived cohort, and its absence here should not be read as rarity.

On that last point, here is the complementary measurement I have, limitation first. My frame is published code, not deployed systems, and I let it be over-quoted as the latter before I corrected the record. The State of MCP Security, 2026 (v1.0, 2026-08-17): 2,303 distinct public MCP server configs from a GitHub crawl plus the official MCP Registry, deduped by content, of which 1,201 (52.1%) declare a remote server with no authentication. Almost none of those 1,201 will ever be a CVE, because a default is not a code defect. Your 24 and my 1,201 are measuring disjoint things, and neither is prevalence. I think both cards should say so.

And I should disclose that I am asking for a validation-provenance field I do not publish myself. My own scanner ships a benign-slice badge reading 536 MCP configs scanned with 0 of 1 HIGH/CRITICAL findings adjudicated a false positive. One adjudication, printed as a percentage. I have no standing here except having made this exact mistake first.

Interested-party disclosure: I maintain agent-audit-kit and agent-airlock, which are static and runtime tools in this space, so a clearer validation boundary in public MCP vulnerability data is useful to me commercially. I have not reproduced your filtering, and I am not disputing any of your figures — only asking which artifact they attach to.

Happy to contribute the field definition and a backfill script for the provenance column if that is useful, and happy to be told the distinction is already documented somewhere I did not read.

Thank you for the helpful suggestions. We agree that the distinction between the paper’s snapshot and the expanded release should be clearer.

  1. Reported figures: The 96.7% agreement applies to the historical sampled exclusion audit; 87.5% and 66.7% describe the original 24-record snapshot. None describes the 791-record release. We will clarify this in the card’s opening paragraph.

  2. Validation provenance: We will add review-provenance metadata, distinguish relevance review from annotation review, and mark unavailable historical provenance as unknown.

  3. Sampling limitations: Agreed—the distributions describe included CVE disclosures, not prevalence across the MCP ecosystem. We will explicitly note that configuration weaknesses and issues without CVEs may be underrepresented.

  4. Ongoing maintenance: We maintain this dataset in our spare time, so human review may progress slowly. We will publish improvements incrementally and continue adding newly disclosed CVEs, tentatively on a monthly schedule.

We will also disclose that the original individual-review records behind the historical agreement figure are no longer available for re-audit. Thank you again for helping us improve the dataset and for offering to contribute.

Sign up or log in to comment