Skill flagged — suspicious patterns detected
ClawHub Security flagged this skill as suspicious. Review the scan results before using.
PinchBench
v1.0.0Run PinchBench benchmarks to evaluate OpenClaw agent performance across real-world tasks. Use when testing model capabilities, comparing models, submitting b...
⭐ 0· 700·3 current·3 all-time
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Suspicious
medium confidencePurpose & Capability
The skill's name/description align with the included code: it loads tasks, creates/runs OpenClaw agents, grades runs, and can upload results to a leaderboard. Minor mismatch: pyproject.toml lists dependencies like fabric and paramiko (SSH-related) that are not mentioned in SKILL.md; their presence may be legitimate for some tasks but is not explained in the README or SKILL.md.
Instruction Scope
Runtime code reads OpenClaw agent configuration and session transcripts from the user's home (~/.openclaw/agents/*) and will prepare agent workspaces (possibly writing files into agent workspaces). The grading engine executes automated grading code via exec() extracted from task markdown — this executes arbitrary Python from task files. The skill also supports registering/uploading results to pinchbench.com, which could transmit transcripts or workspace contents to an external server.
Install Mechanism
There is no install spec (instruction-only), so nothing downloads or runs during installation. Code files are included in the skill bundle. The project metadata (pyproject.toml) lists third-party dependencies (pyyaml, fabric, paramiko) that may need to be installed if the user runs the scripts; SKILL.md only mentions 'uv' and Python. No remote download URLs or extract steps were found.
Credentials
The skill declares no required environment variables but reads local OpenClaw state (workspaces, sessions) and may create agents via the openclaw CLI. It can register an API token and upload results to a public leaderboard (pinchbench.com). Requesting no env vars while accessing local agent data and offering an upload path is proportionate to benchmarking, but the lack of explicit warning about what is uploaded (transcripts, workspace files) is a concern.
Persistence & Privilege
always is false and the skill does not request elevated platform privileges. At runtime it will create OpenClaw agents (via the openclaw CLI), create workspaces (in ~/.openclaw or /tmp), and may write a stored token/config when --register is used. Creating agent entries and writing token/config files is coherent with its purpose but users should expect persistent artifacts under their OpenClaw config and /tmp.
What to consider before installing
What to check before installing/using this skill:
- Inspect lib_upload.py (and the --register/upload code path) to see exactly what fields are sent to pinchbench.com; do not upload runs containing private data unless you understand what's transmitted. Use --no-upload for local-only testing.
- Review the grading code behavior: tasks may contain embedded Python automated checks which the grading engine executes with exec(); only run tasks from trusted sources or review task_*.md files before running.
- Be aware the skill reads OpenClaw files in your home (~/.openclaw/agents/*) and can create new agents and workspaces via the openclaw CLI — this may expose agent transcripts or sensitive tool outputs to local processing and (if you upload) to the leaderboard.
- The pyproject lists fabric/paramiko (SSH capabilities) not mentioned in the docs — if you install dependencies, consider doing so in an isolated virtualenv or sandbox and review why SSH libraries are needed.
- If you want to be cautious: run the scripts in an isolated environment (container or VM), run with --no-upload first, and audit any results JSON for sensitive content before sharing. If you plan to allow uploading, confirm the upload endpoint and privacy policy on pinchbench.com and inspect where tokens are stored by the registration flow.Like a lobster shell, security has layers — review code before you run it.
latestvk97cr769b6epjg1bg1cmhv2acx81b1mr
License
MIT-0
Free to use, modify, and redistribute. No attribution required.