The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates an application pattern that ingests untrusted text from external web pages and maps it directly into language model contexts, opening an exploitation surface for indirect prompt injection.
Ingestion points: File SKILL.md contains the ArchiveWorkflow.process_url method which retrieves external data via self.scraper.fetch(url).
Boundary markers: Although the markdown documentation outlines an untrusted content contract, the actual code implementations of the prompts in ArchiveCategorizer.categorize and EntityExtractor.extract lack structural boundary delimiters or instructions to ignore embedded commands.
Capability inventory: The resulting enriched data structures are dynamically parsed into JSON, written to Google Sheets via SheetsService, and compiled into local files via ArchivePDFGenerator.
Sanitization: Content is truncated via slicing ([:8000] and [:10000]), but no validation or character escaping is carried out to sanitize potential adversarial prompts inside the content.

digital-archive