The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a 'Stage Output Correction' loop in SKILL.md that automatically re-feeds malformed output back into the AI's prompt for correction. If a subagent's output is manipulated by malicious instructions within an artifact, this mechanism could unintentionally repeat and potentially execute those instructions during the retry attempt.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted 'original artifacts' across multiple stages without sufficient isolation.
Ingestion points: Untrusted data is ingested as 'original artifacts' in SKILL.md and passed to the subagents defined in references/initializer.md, references/normalizer.md, references/adversary.md, and references/referee.md.
Boundary markers: The subagent prompt templates do not use clear delimiters or explicit instructions to treat the artifact content as data rather than instructions.
Capability inventory: The skill has the ability to persist data to the .context/triangulate/ directory and generate a final findings report, which could be used to store or display malicious output.
Sanitization: There is no evidence of sanitization or validation of the input artifacts to strip potential injection strings before processing.

triangulate