The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the osv-scanner binary from Google's official GitHub repository and installs Python security tools (semgrep, slither-analyzer) via pip at runtime.
[REMOTE_CODE_EXECUTION]: During the fuzzing phase (Arm A3.5) and the PoC verification gate (Arm A4.5), the skill compiles and executes code from the untrusted third-party repositories it audits. While it includes environment scrubbing (using env -u) to protect its own secrets, it still executes arbitrary code from external sources.
[COMMAND_EXECUTION]: The skill uses chmod +x to make downloaded binaries executable and performs various shell operations including gh api, cargo fuzz, and local reproduction scripts against cloned repositories.
[DATA_EXFILTRATION]: The disclose arm (Arm C) is designed to automatically send disclosure emails to third-party maintainers using a Resend API key. While it includes multiple safety gates (budget, cooldown, and arming flags), it represents an automated network channel for sending data externally.
[INDIRECT_PROMPT_INJECTION]: The skill audits untrusted codebases, reading files like README.md, SECURITY.md, and source code. This represents an attack surface where a malicious repository could include instructions designed to influence the agent during its 'Agentic logic audit' or triage phases.
Ingestion points: git clone of arbitrary repositories, gh api calls to fetch repository metadata and security policies.
Boundary markers: The skill employs secret scrubbing (env -u) before executing target code and uses budget limits (N entrypoints) to bound agentic reviews.
Capability inventory: Access to GitHub API, Resend email API, package managers (pip, npm, cargo), file system modification (chmod), and network operations (curl, secretcurl).
Sanitization: Implements redacted logging for PoC verification and scrubs Aeon credentials from the environment when running target-controlled code.
[DYNAMIC_EXECUTION]: The skill involves runtime compilation and execution of external code from audited repositories, specifically using cargo fuzz for Rust projects and Foundry/local scripts for vulnerability verification.

vuln-scanner