agent-scan

NVIDIA SkillSpector examines Git repositories, URLs, archives, directories, or individual files before an agent skill is installed. It checks dozens of patterns spanning prompt injection, exfiltration, privilege escalation, dangerous code, dependency vulnerabilities, MCP least privilege, and tool poisoning, then produces terminal, JSON, Markdown, or SARIF reports with risk scores. Scanned skills are never executed. Optional LLM analysis sends file contents to the configured provider, while static-only mode keeps that content local.

Semia analyzes an Agent Skill before the user decides to trust or run it. The scanner reads the skill as data and never executes its embedded commands, then identifies possible actions, effects, network access, secret reads, and tool invocations with evidence tied to source lines. It can be used through a Python CLI or installed as a plugin for Codex, Claude Code, or OpenClaw. A configured LLM provider is required for the full prepare, synthesize, detect, and report workflow.

This command-line scanner reviews third-party Claude Skills before installation for suspicious network access, sensitive file operations, dangerous commands, dynamic execution, and injection patterns. It can produce HTML, console, or JSON reports and supports custom rules and allowlists, while performing static analysis without executing the skill code. Its findings are a review aid rather than a guarantee that every security risk will be detected.

AgentDiff captures AI-assisted edits across Claude Code, Cursor, Copilot, Codex, Windsurf, OpenCode, and Gemini, signs line-level attribution with Ed25519, and stores the evidence in Git. Its blame, diff, context, and report commands show who changed what. Pushed traces can expose short prompt excerpts to repository readers, so sensitive teams may need to disable prompt capture.