The Agent Skills Directory

[PROMPT_INJECTION]: The generated .githooks/pre-push script is vulnerable to Indirect Prompt Injection. It passes untrusted data (git diffs and commit messages) directly into an AI agent's prompt.
[COMMAND_EXECUTION]: The pre-push hook executes the claude CLI using the --dangerously-skip-permissions flag. This flag is designed to allow the agent to perform actions like running shell commands and modifying files without user confirmation. When combined with untrusted input from a git diff, an attacker could include malicious instructions in a codebase that the hook-triggered agent would execute with high autonomy.
[COMMAND_EXECUTION]: The justfile and Phase 3 setup instructions execute commands that download and run code from external registries (npm, npx, uv). While these are common developer tools, the automated nature of the scaffolded scripts and hooks increases the attack surface.

project-init