The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents dasel v3, a tool designed to process untrusted structured data (JSON, YAML, CSV, etc.) from external sources. The tool provides powerful functions such as readFile(path) and parse(format, content) which can be triggered by selectors. If an agent processes data where the selector is partially controlled by an attacker or the data contains self-referential instructions, it could lead to unauthorized file reads.
Ingestion points: dasel processes data from stdin or piped files as documented in SKILL.md and references/format-patterns.md.
Boundary markers: No specific safety delimiters or warnings regarding the handling of untrusted input are mentioned in the reference.
Capability inventory: Includes readFile (filesystem access), parse (dynamic data parsing), and base64d (decoding) which can be combined to process and execute complex data-driven logic.
Sanitization: The reference does not discuss sanitization of input data before processing.
[DATA_EXFILTRATION]: The skill documents the readFile(path) function in references/functions.md, which allows the tool to read arbitrary files from the local filesystem. While documented as a standard feature of the dasel CLI, this capability provides a primitive for data exposure if used by an agent to access sensitive configuration or credential files on the host system.

dasel-reference