The Agent Skills Directory

[COMMAND_EXECUTION]: The provided CI/CD integration examples for GitHub Actions and GitLab CI (found in SKILL.md and the references/ directory) contain command injection vulnerabilities. User-supplied inputs, such as inputs.prompt, inputs.model, and inputs.prompts_file, are directly interpolated into shell command strings (e.g., in run blocks) without proper sanitization or the use of environment variables. An attacker could exploit this by providing inputs containing shell metacharacters to execute arbitrary commands on the CI runner.\n- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and processing untrusted data from CI inputs and external configuration files.\n
Ingestion points: Untrusted data is read from workflow_dispatch inputs and external files like video-prompts.yml and prompts/weekly-content.json.\n
Boundary markers: There are no structural boundary markers or delimiters implemented in the provided code to isolate instructions from untrusted data, although the documentation mentions the importance of validation.\n
Capability inventory: The skill uses the requests library to communicate with the Kling AI API and performing file system writes to save generated videos and metadata to local directories.\n
Sanitization: The implementation lacks input validation and sanitization mechanisms, which could allow malicious instructions embedded within prompts or manifests to influence the skill's operations.

klingai-ci-integration