The Agent Skills Directory

REMOTE_CODE_EXECUTION (HIGH): The agent instruction file agents/rust-project-init.md contains a command to install the beads tool using curl -sSL [URL] | bash. This pattern is high-risk as it executes unverified code from an external repository (steveyegge/beads) directly in the shell without integrity checks.
EXTERNAL_DOWNLOADS (MEDIUM): Multiple files (e.g., skills/backend-rust/scripts/audit.py, agents/rust-project-init.md, skills/secrets-guardian/SKILL.md) automatically trigger package installations via pip install, cargo install, and npm install for various tools. While these tools are generally known (e.g., pre-commit, cargo-audit), unversioned installations from public registries pose a supply chain risk.
COMMAND_EXECUTION (LOW): The skill makes extensive use of the Bash tool to run local build processes, start development servers, and execute tests. This behavior is dropped to LOW severity as it is strictly necessary for the primary purpose of the skill (coding and local development).
PROMPT_INJECTION (LOW): The skill has a high surface area for indirect prompt injection. It ingests untrusted user "ideas" and processes them using highly capable agents (vibe-coder, rust-developer) that have Write, Edit, and Bash permissions.
Ingestion points: User-provided idea strings in commands/ship.md and agents/vibe-coder.md.
Boundary markers: Missing; the user's natural language idea is directly used to guide code generation without explicit delimiters or "ignore embedded instructions" warnings.
Capability inventory: Full shell access (Bash) and file system access (Write, Edit) across all developer agents.
Sanitization: None detected for the initial input processing phase.

vibe-coder