ClawVet is a browser-based security scanner for AI-agent skills and related files, including SKILL.md, Cursor rules, AGENTS.md, and helper scripts. Users can paste or fetch a file and inspect what its code may access before running it. The scanner works entirely on the device using WebAssembly and Tree-sitter, combining static rules with AST-based taint tracking to identify exposed secrets, contacted hosts, startup-file changes, and other permissions. It assigns a risk grade and severity-ranked findings, with lines users can inspect. A command-line version scans multi-file bundles; installation-time checks and a GitHub Action can flag permission changes, helping detect when an update quietly expands a skill’s capabilities.
The project reports results from SkillTrustBench v1.1, a held-out set of 556 skills: ClawVet scored 0.833 accuracy and 0.770 macro F1, compared with lower reported scores for Cisco and NVIDIA scanners. Its rug-pull detection caught 95.2% of code-injection updates, while prompting on 5.1% of benign updates; this detection uses no language model. A separate audit of 209 ClawHub skills found that 40.2% read undeclared secret environment variables and 4.8% accessed host credential stores. These figures frame the tool’s purpose: revealing risky behavior and permission changes that may be easy to miss when reviewing agent skills.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.