The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The scripts/runner_recon.sh script is designed to search for and identify sensitive credential files such as ~/.git-credentials, ~/.netrc, and private SSH keys within the ~/.ssh/ directory.
[REMOTE_CODE_EXECUTION]: Documentation within the skill, specifically references/pipeline-poisoning.md, includes functional examples of remote code execution payloads, such as a bash reverse shell string, for vulnerability demonstration.
[COMMAND_EXECUTION]: The skill utilizes bash scripts (runner_recon.sh, provenance_verify.sh) to execute various system and container utilities, including ps, hostname, id, ls, docker, and crane.
[EXTERNAL_DOWNLOADS]: Several Python scripts utilize the requests library to fetch data from external sources, including the GitHub API and public package registries such as NPM and PyPI, to facilitate security auditing.
[DATA_EXFILTRATION]: The skill provides detailed methodologies and example payloads for exfiltrating CI/CD secrets to external out-of-band domains, often using encoding techniques to bypass standard logging filters.
[PROMPT_INJECTION]: The skill operates as an audit tool that ingests and processes untrusted pipeline configuration files.
Ingestion points: The scripts/workflow_auditor.py and scripts/malicious_action_scanner.py scripts read local workflow YAML files and project manifests.
Boundary markers: The standalone scripts do not utilize specific delimiters to isolate untrusted data during analysis.
Capability inventory: The skill possesses capabilities for file system read operations, network communication via HTTP, and subprocess execution.
Sanitization: Secure parsing of YAML files is implemented using yaml.safe_load in scripts/workflow_auditor.py.

cicd-supply-chain