The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from user-provided file paths, URLs, and pasted text. It lacks specific instructions to treat the ingested content strictly as data, potentially allowing embedded malicious instructions to influence the agent's behavior during the review process.
Ingestion points: The 'Input Handling' section in SKILL.md specifies reading from file paths, fetching from URLs, and processing pasted content.
Boundary markers: No delimiters or explicit 'ignore embedded instructions' warnings are provided for the document analysis phase.
Capability inventory: The skill utilizes file reading, network fetching, and shell command execution.
Sanitization: No sanitization or validation of the ingested content is specified before the agent performs the adversarial review.
[COMMAND_EXECUTION]: The skill instructs the agent to use CLI tools to process specific file types, specifically using shell commands for document conversion.
Evidence: SKILL.md specifies: 'download it with curl and extract the text -- prefer pdftotext if available'.
[EXTERNAL_DOWNLOADS]: The skill functionality involves fetching content from arbitrary external URLs provided by the user.
Evidence: SKILL.md instructs: 'Use WebFetch to download the content... If the URL points to a PDF, download it with curl'.

ape-poke-holes