The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to execute various shell commands, including git, gh, grep, and node, for repository analysis and diff inspection.- [REMOTE_CODE_EXECUTION]: Section §7 contains a node -e one-liner that encourages the agent to paste code segments (the Sequelize where clause) from the PR diff into the execution string. This dynamic code generation from untrusted diff content facilitates a potential command injection attack, as a malicious PR could break out of the JavaScript context to run arbitrary system commands.- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to processing untrusted data from PR titles, bodies, and logs. 1. Ingestion points: git diff, git log, gh pr view. 2. Boundary markers: Absent. 3. Capability inventory: Shell access and dynamic JavaScript execution. 4. Sanitization: Absent.- [DATA_EXPOSURE]: The skill uses grep to identify sensitive information like secrets and PII in the codebase. Although intended for auditing, this demonstrates the ability to read and process sensitive environmental data.

cx-pr-review