The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes and summarizes external project code to update research artifacts. This ingestion of untrusted data allows for potential indirect prompt injection if the source code contains malicious instructions designed to alter the agent's behavior during its analysis.
Ingestion points: Reads existing research artifact files and external source code via tools like Read, Grep, and the artifact.mjs helper script (e.g., in SKILL.md steps 1 and 2).
Boundary markers: The artifacts use a structured format with Markdown frontmatter (---) and HTML-style comment tags (), providing structural separation, though not explicitly preventing the LLM from following instructions within the content.
Capability inventory: The skill uses Bash, Read, Write, Grep, and Glob tools. The artifact.mjs script can read, write, and rename files on the filesystem using node:fs/promises.
Sanitization: The artifact.mjs script performs sanitization on metadata and comment attributes (using oneLine and attr functions) to prevent the injection of structural tags that could corrupt the artifact format. However, it does not sanitize the primary research findings derived from analyzed code.

iterate-research