harnesswatch: data

The structured data behind harnesswatch, an open-source security audit toolkit and the study The Harness Is the Attack Surface: A Security Analysis of AI Coding-Agent Harnesses (2025–2026). The harness is the runtime loop around the model (Claude Code, Codex CLI, Gemini CLI, GitHub Copilot, Cursor, Kiro, Grok Build, Antigravity, Goose, Aider and others) that holds credentials, runs shell commands, starts MCP servers, reads repository configuration and runs unattended in CI.

• Paper (web): https://thanatchon36.github.io/harnesswatch/paper/
• Code and documentation: https://github.com/thanatchon36/harnesswatch
• Archive (all versions): https://doi.org/10.5281/zenodo.23407630

How this data was made. LLM research agents (Claude, operated through Claude Code) collected the evidence, produced both matrix ratings and collected and classified the incidents, working from the author's written specification. The paper's Method section gives the full disclosure. Claude Code is itself among the systems studied. Rows carry a source URL, evidence or method so that they can be re-checked. Closed-source harnesses were studied black-box only (no decompilation, no leaked source). Benchmark payloads are inert and only synthetic secrets were used.

Control matrix (RQ2): 0–3 score per dimension for 58 harnesses, 664 evidence-linked cells.

Verified harness incidents and CVEs by quarter of disclosure (RQ3); the final quarter is partial.

Configs

Config	Rows	What it is
incidents (default)	357	RQ3 table of harness-level advisories, CVEs, design weaknesses and incidents, with taxonomy class, root-cause class, affected and fixed versions, patch latency and primary/secondary sources. The paper counts 323 of these rows. It excludes 29 unverified rows, 4 that failed the independent CVE/GHSA check and 1 disclosed outside 2025–2026. There is no counted column: experiments/rq3_incidents.py in the repository applies the rule.
ledger	1519	Every factual claim used in the project: claim text, harness, evidence label (verified-by-test, from-official-docs, third-party-report, unverified), source URL, publisher, access date, verifier and confidence.
controls	880	RQ2 control matrix, rater 1: one row per harness × dimension with score 0–3 or an N/A reason, rationale and evidence links. The paper uses D01–D15, which gives 664 scored cells and 206 N/A. Ten further rows are D16, which is outside the paper's fifteen-dimension rubric, so filter them out to reproduce its figures. A blind second LLM-agent rater working from the same fact lists reached quadratic κ = 0.88 (raw/experiments/e1_agreement.json).
claims_vs_observed	117	RQ7: documented vendor claims for closed harnesses against what a black-box probe observed, with method, date and verdict. 28 claims were testable headless; 22 matched.
incident_classification	328	Taxonomy and root-cause label per incident with a one-line rationale.
bench_runs	350	RQ5 canary prompt-injection benchmark: 7 harnesses × 2 permission modes × 5 scenarios (one control, four injection) × 5 repetitions on one small local model, of which the paper analyses 340. One row per run, with injection-followed, exfiltration, destructive-action, secret-in-state, permission-prompt and task-completed flags, hosts contacted and timing. List fields are JSON strings.

Raw files

raw/ holds the files unchanged from the repository, for anyone who prefers the originals. The column reference in raw/evidence/README.md predates the final taxonomy, so raw/taxonomy/ is authoritative for class definitions:

• raw/evidence/: incidents.csv, ledger.csv, sources.bib (BibTeX for every cited source, with urldate) and the column reference.
• raw/matrix/: controls.csv, claims_vs_observed.csv, the scoring rubric (rubric.md, rubric.yaml) and the dimension definitions.
• raw/taxonomy/: the threat taxonomy (harness-threats.yaml, T01–T19 mapped to STRIDE, MITRE ATLAS, OWASP LLM/Agentic/MCP), root-causes.yaml and incident-classification.csv.
• raw/bench/runs.jsonl: the benchmark runs as recorded (per-run metadata and outcome flags, no request bodies).
• raw/experiments/: the versioned results JSON from which every number in the paper is generated (E1 matrix and agreement, E2 prevalence corpus and strata, E3 benchmark, E4 supply chain, E5 secret residue, E6 detector accuracy, E7 claims, RQ3 incidents).

Loading

from datasets import load_dataset

incidents = load_dataset("thanatchon36/harnesswatch", "incidents", split="train")
controls  = load_dataset("thanatchon36/harnesswatch", "controls",  split="train")
runs      = load_dataset("thanatchon36/harnesswatch", "bench_runs", split="train")

Study snapshot

Study cut-off 15 October 2026. Harness versions are recorded in the controls and claims_vs_observed rows and in the repository's VERSIONS.md.

Ethics and intended use

Defensive and research use. The incident rows describe publicly disclosed issues and cite their public sources; nothing here is an undisclosed vulnerability. The benchmark rows record the behaviour of harnesses on inert canary prompts in sandboxed containers with egress blocked. Do not read the benchmark as a ranking of harness resilience: with five runs per scenario cell on one small model, the intervals are wide and the paper draws no such ranking.

Citation

@misc{chongmankhong2026harness,
  title  = {The Harness Is the Attack Surface: A Security Analysis of AI Coding-Agent Harnesses (2025--2026)},
  author = {Chongmankhong, Thanatchon},
  year   = {2026},
  doi    = {10.5281/zenodo.23407630},
  url    = {https://thanatchon36.github.io/harnesswatch/paper/}
}

License: Apache-2.0, the same as the repository.