The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data via the $ARGUMENTS variable in the Requirements section and processes external code diffs and PR descriptions. These inputs are interpolated directly into LLM prompts (e.g., the review_prompt in the Python example) without robust boundary markers or sanitization. This creates a surface for indirect prompt injection, where an attacker could embed malicious instructions within a pull request to manipulate the AI's analysis results.
[COMMAND_EXECUTION]: The provided code snippets and CI/CD templates utilize subprocess.run, subprocess.check_output, and shell pipes to execute static analysis tools like sonar-scanner, semgrep, trufflehog, and jq. While these are intended for legitimate security and quality analysis, the use of external arguments (such as --repo in the Python script) without explicit validation represents a potential command injection vector if the script is executed in an automated environment with untrusted inputs.

code-review-ai-ai-review