The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from Jenkins, including build logs, job descriptions, and configuration XML. Because the skill also has the capability to perform network POST requests (triggering builds) and modifying configurations, this creates a potential surface for indirect prompt injection.
Ingestion points: Jenkins API responses for logs, metadata, and job configurations (references/remote-access-api.md).
Boundary markers: The instructions mention providing "desensitized summaries" but do not define explicit delimiters or escaping for the ingested content.
Capability inventory: Network POST requests via curl, filesystem writes to temporary directories.
Sanitization: The skill includes logic to quote credentials and mask sensitive parameters, but does not explicitly sanitize log content against adversarial instructions.
[DYNAMIC_EXECUTION]: The skill uses inline Python scripts executed via stdin to process credentials and generate configuration files at runtime.
Evidence: The reference file references/remote-access-api.md uses `python3
<<'PY'` to parse JSON credentials and format curl config files.
Mitigation: The skill implements robust defenses by using mktemp for directory creation, umask 077, and chmod 600 for temporary configuration files, and trap for cleanup on exit.
[COMMAND_EXECUTION]: The skill uses system binaries curl and python3 to perform its primary functions.
Evidence: Requirement specified in SKILL.md metadata and usage throughout references/remote-access-api.md.

jenkins