Clawtributor Security Reporter
Prepare sanitized AI-agent incident reports for explicit human approval before submission.
The evidence shows local-first report storage, explicit approval for every submission, sanitization guidance, and disclosure that reports are sent to a public GitHub form; the permission scope is relatively narrow. Points are deducted because installation relies on remote npx/ClawHub sources, rollback and cleanup are not specified, and signing verification depends on externally downloaded artifacts and local tooling with limited post-egress governance.
The documentation provides a report JSON schema, state-file example, manual submission steps, and a privacy checklist, making the normal path reasonably clear. Static review found no clawtributor-specific tests or evidence for abnormal-input handling and failure feedback; the schemas differ between SKILL.md and reporting.md, so the score remains below the static ceiling of 10.
The audience, supported platforms, trigger phrases, report types, and manual-submission boundary are described. Points are deducted because there is no Chinese documentation or example, platform behavior is documented mainly through English POSIX commands, and the core submission path depends on a GitHub browser form that may be difficult to reach from mainland China.
The skill includes a name, version 0.0.9, license, homepage, skill.json, changelog, installation guidance, examples, and Prompt Security attribution. Points are deducted for duplicated documentation and inconsistent report schemas, dependencies not being consolidated, and the absence of a clearly named maintainer, compatibility policy, troubleshooting guidance, or rollback procedure.
The skill can directly guide an agent to create a standardized security incident report, save a local draft, obtain user confirmation, and submit it manually; the core goal is substantially covered. Points are deducted because final submission still requires user-operated browser copy/paste, no implementation code or skill-specific output verification is provided, and static calibration limits the score to 7.
The evidence includes concrete report schemas, a state example, signed-release verification instructions, and versioned changelog entries, making some claims auditable. Points are deducted because nothing was executed, repository CI evidence does not cover clawtributor's key paths, and there is no independent reproduction or corroborating third-party evidence; static review therefore stays below the ceiling of 5.
- Review and sanitize every report before approval; submitted reports leave the host and enter a public GitHub issue workflow.
- Installation and artifact verification depend on external services or tools including npx, ClawHub, GitHub, curl, jq, and OpenSSL; verify source reachability and tool availability first.
- The report fields differ between SKILL.md and reporting.md; select one schema and validate it before submission.
- Recovery, local-state cleanup, and withdrawal of an already submitted report are not documented.
What does this skill do, and when should you use it?
Clawtributor is a community security incident reporting skill for OpenClaw, NanoClaw, Hermes, and Picoclaw. It helps agents document malicious prompts, vulnerable skills or plugins, and tampering with security tooling. Reports remain local by default and are submitted manually through a browser only after explicit user approval. The skill does not automatically submit reports and requires sensitive evidence to be sanitized before sharing.
It prepares JSON reports containing the report type, severity, title, description, evidence, affected skill, and recommended action; stores reports under ~/.clawsec/clawtributor/reports/ with private permissions recommended via chmod 600; shows the exact proposed submission and requests explicit approval; after approval, directs the user to the GitHub Security Incident Report form for manual submission; and tracks submissions in ~/.clawsec/clawtributor/state.json. For standalone installation, it also provides commands using curl, OpenSSL, jq, and SHA-256 tools to verify the signed release manifest, archive, SKILL.md, and skill.json.
- A security engineer can document a prompt-injection attempt that tried to bypass controls, extract secrets, or override user intent.
- A platform maintainer can prepare a vulnerability report when a skill or plugin exfiltrates data, requests unjustified permissions, or attempts self-modification.
- An agent operator can record suspected tampering with security files, audit jobs, advisory URLs, or health checks.
- A team can contribute to ClawSec community threat intelligence while retaining local control over evidence and approving each submission individually.
What are this skill's strengths and limitations?
- Covers malicious prompts, vulnerable skills or plugins, and security-tool tampering.
- Keeps reports local by default and gates every submission on explicit approval.
- Includes evidence-sanitization guidance and local state tracking.
- Supports four target agent platforms and documents signed release-artifact verification.
- It does not submit automatically; approved reports still require manual browser-based GitHub submission.
- Runtime networking is disabled by default, but standalone release verification fetches GitHub assets.
- The supplied SKILL.md does not document automatic discovery, automatic sanitization, or a test suite.
- It relies on local shell and filesystem access plus npx, curl, OpenSSL, and jq.
How do you install this skill?
For a Vercel Skills-compatible installer, run: npx skills add prompt-security/clawsec --skill clawtributor -a codex -y. In OpenClaw/ClawHub environments, run: npx clawhub@latest install clawtributor. For a standalone release, run the verification procedure in SKILL.md before installing or extracting the archive.
How do you use this skill?
Ask the agent to prepare a report, for example: “Prepare a sanitized report for this suspected prompt injection, remove personal data and credentials, keep it local, and do not submit it.” Reports should be saved under ~/.clawsec/clawtributor/reports/. Before submission, the agent must show the exact report type, severity, title, and summary and ask for approval. Only after explicit approval should it open the GitHub Security Incident Report form for manual submission.