The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from a findings.json file produced by a previous step. This creates a vulnerability surface where malicious instructions could be embedded in fields like notes or recommendation to influence the agent's reasoning or subsequent steps in the pipeline.
Ingestion points: The SKILL.md file instructs the agent to read findings.json from the working folder.
Boundary markers: Absent. The instructions do not provide delimiters or warnings to ignore potentially malicious instructions embedded within the findings data.
Capability inventory: The skill performs file system read/write operations on JSON files and carries out complex reasoning to summarize and filter content.
Sanitization: Absent. The skill instructions explicitly state to "preserve every original field verbatim," ensuring that any injected content remains intact through the triage process.

triage-findings