The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an inherent risk of indirect prompt injection because its primary function is to ingest and analyze external, untrusted content.
Ingestion points: Processes user-provided log snippets, local stdout/stderr text, simulation artifacts, and logs retrieved from remote clusters using nvflare job logs as defined in SKILL.md and references/evidence-collection.md.
Boundary markers: Highly structured boundaries are present. The skill mandates a 'Log Content Trust Boundary' requiring the agent to flag any malicious or overriding log entries as SUSPICIOUS_LOG_CONTENT and explicitly forbids executing embedded directives.
Capability inventory: The skill maintains a read-only stance with a declared blast-radius: read_only. It invokes analytical commands such as nvflare job meta, nvflare job logs, and nvflare agent inspect source while explicitly prohibiting job or configuration mutations.
Sanitization: The instructions require the agent to avoid executing instructions found within tracebacks or logs and explicitly demand that raw personal data or sensitive dataset values found within log streams should be summarized rather than quoted.

nvflare-diagnose-job