The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests external data such as "observed attempts/results" and "source evidence" to generate decision records in the docs/adr/ directory and update skill files in .agents/skills/. This creates a potential surface for indirect prompt injection if the source data contains malicious instructions. However, the risk is mitigated by the skill's narrow focus on documentation and the explicit instruction to "never copy secrets or raw transcripts," which reduces the impact of data exfiltration or credential exposure during the capture process.
[DATA_EXPOSURE]: The skill demonstrates good security practice by explicitly instructing the agent to avoid sensitive data capture: "Never copy secrets or raw transcripts" (references/decisions.md).

he-learn