caplock-bench / README.md
NagaYu's picture
CapLock Bench v0 (dev split)
121fdb1 verified
|
Raw History Blame Contribute Delete
3.72 kB
metadata
license: other
license_name: all-rights-reserved
license_link: LICENSE
pretty_name: CapLock Bench v0 (dev split)
tags:
  - mcp
  - model-context-protocol
  - supply-chain
  - security
  - synthetic
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: dev
        path: data/dev.jsonl

CapLock Bench v0 — dev split

A small, synthetic benchmark of changes to an approved MCP (Model Context Protocol) server setup: poisoned tool descriptions, schema expansion, annotation flips, tool additions/removals, package integrity substitution (npm / PyPI / docker), floating version specs, and benign controls (key order, formatting, Unicode normalization, number spelling, version-only bumps).

This repository contains the dev split only (42 cases, 21 operator families). The holdout split (48 cases) is withheld so that it stays usable as an evaluation set: its cases are not published here or in the source repository. Its identity can be checked on any future disclosure against corpusDigest in the repository's bench/results.json.

Source repository: https://github.com/NagaYu/caplock

What a row is

field meaning
id, family, axis, seed case id, mutation operator family, axis, and the scenario it generalizes
base the synthetic starting workspace (demo: notes over npx + weather over uvx; docker)
mutation_json JSON: replacement tools/list documents, config file text, or registry metadata overlays
expected_codes finding codes (CapLock rule catalog 1.0.0) that describe the change
expected_max_severity the threat class a reviewer should be alerted at (unverifiable, critical, high, medium, info, or null)
expected_exit exit code under --fail-on high (0 pass, 1 findings, 2 unverifiable)
benign true for controls that should not raise anything beyond version_changed

To replay a case you need the CapLock repository (the base workspace and the synthetic fixture server); the row alone describes the mutation, not the full environment.

How it was made

  • Generated by bench/make_corpus.py from hand-written operators; labels were written from each operator's meaning and committed before the first scoring run, and not edited afterwards.
  • All packages (@caplock-demo/*, caplock-demo-weather, ghcr.io/caplock-demo/*) and tool definitions are synthetic. No real MCP server's tool definitions are included.
  • Whole operator families are assigned to one split, so holdout contains patterns not present here.

Reference numbers (CapLock 0.1.0 on this split)

Detection 36/36 (100.0%), false positives on benign controls 0/6 (0.0%), exit-code agreement 42/42. The dev split is expected to be easy for CapLock: the operators were written alongside its rules. See the holdout numbers and the limits in the CapLock repository's bench/RESULTS.md.

Limitations

  • Synthetic and small (two fixture servers, four tools); real servers are larger and messier.
  • Self-authored operators and labels: over-fitting to the author's own detector is likely.
  • expected_codes follow CapLock's taxonomy; other tools will name things differently.
  • Nothing here measures runtime behavior. CapLock pins the declared tool surface and the package identity. It does not attest to the runtime behavior or safety of any server.

License and contact

All Rights Reserved — Copyright (c) 2026 Yuta Nagao (see LICENSE). This dataset is public for evaluation purposes only; viewing it does not grant any licence to use, copy, modify, redistribute, or train on it. The project is available for acquisition or licensing — contact via https://nagayu-portfolio.nagayu.workers.dev/en