report-clawhub-malicious-skill

name	report-clawhub-malicious-skill
description	Use when a researcher, maintainer, or contributor found or suspects a malicious skill on ClawHub and needs a private reporting workflow: opening a GitHub private vulnerability report, preparing a proposal-only PR, creating `proposals/<GHSA-ID>/clawscan.yml`, running ClawScan against the malicious skill, and explaining what evidence belongs in private versus public channels.

Report ClawHub Malicious Skill

Overview

Use this workflow to help someone report a malicious ClawHub skill and propose a ClawScan profile change that catches it. Keep sensitive details private and make the public PR contain only the candidate ClawScan config.

Safety Boundary

Separate private evidence from public contribution material:

• Put live malicious skill URLs/slugs, impact, exploit details, reproduction notes, and private ClawScan artifacts in GitHub private vulnerability reporting.
• Do not paste secrets, exploit payloads, private artifacts, or live bypass details into public issues, public PR text, README edits, or config comments.
• Do not ask the reporter to execute the suspicious skill. ClawScan should scan a local copy as data; it should not run the skill's behavior.
• If the reporter does not have a GHSA/private report id yet, have them open the private report first and wait for the identifier before opening the public proposal PR.

Workflow

1. Open a GitHub private vulnerability report.

   Ask the reporter to include:

   • affected ClawHub skill URLs or slugs
   • why the current clawhub profile missed it
   • why the proposed ClawScan config catches it
   • local ClawScan artifact paths or attached artifacts
   • any private reproduction context maintainers need

2. Create a proposal-only branch and config.

   The public PR should add only:

   proposals/<GHSA-ID>/clawscan.yml
   

   The file must define a clawhub profile. Start from the current profile and change only what is needed to catch the reported case:

   version: 1
   profiles:
     clawhub:
      scanners:
        - skillspector
        - clawscan-static
       judge:
         command: >
           # candidate judge command, if changed
   

   Do not edit the official bundled profile files in the proposal PR:

   • internal/profiles/clawhub/clawscan.yml
   • internal/profiles/clawhub/prompt.md
   • internal/profiles/clawhub/output.schema.json

3. Prove the candidate catches the reported skill locally.

   Use a local copy of the suspicious skill:

   clawscan /path/to/suspect-skill \
     --config proposals/<GHSA-ID>/clawscan.yml \
     --profile clawhub \
     --output ./artifacts/reported-skill-candidate.json
   

   If useful, compare against the built-in profile:

   clawscan /path/to/suspect-skill \
     --profile clawhub \
     --output ./artifacts/reported-skill-current.json
   

   The public PR should not include the private artifacts unless maintainers say the contents are safe to publish. Reference them in the private report.

4. Open the public PR.

   Keep the PR text minimal:

   • state that a private vulnerability report exists
   • point to proposals/<GHSA-ID>/clawscan.yml
   • do not include sensitive skill details
   • say maintainers should run the official SkillTrustBench Profile Gate

5. Explain maintainer validation.

   Maintainers run the official gate against the proposal:

   clawscan benchmark SkillTrustBench \
     --ids https://huggingface.co/datasets/cuhk-zhuque/SkillTrustBench-results/resolve/main/data/evaluation_subset_10pct.jsonl \
     --config proposals/<GHSA-ID>/clawscan.yml \
     --profile clawhub \
     --output ./artifacts/skilltrustbench-candidate.json
   

   The proposal's clawhub profile shadows the built-in profile for that run. Maintainers review the private report, upload/preserve the benchmark artifact, add a dated baseline under benchmarks/skilltrustbench-leaderboard-10pct/, and port accepted behavior into the bundled clawhub profile. The latest YYYY-MM-DD.json baseline is the current accepted baseline.

Output Shape

When walking someone through the process, end with:

• private report checklist
• exact proposal file path
• candidate config draft or edits needed
• local ClawScan commands to run
• public PR checklist
• clear note that maintainers own the official SkillTrustBench gate and final built-in profile port