The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted text from a local datalake and Wikipedia, which could contain malicious instructions (indirect prompt injection).
Ingestion points: The build_bank.py script reads from /mnt/storage12tb/extractor_corpus/ and the Wikipedia API.
Boundary markers: Output in run.sh uses descriptive headers but does not include explicit delimiters to prevent the agent from following instructions embedded in the chunks.
Capability inventory: The skill has access to Bash, Read, and Write tools.
Sanitization: Content is chunked but not sanitized for potential prompt injection patterns.
[EXTERNAL_DOWNLOADS]: The skill downloads content from a well-known service.
build_bank.py fetches article content from Wikipedia to populate the wikipedia domain chunks.
[COMMAND_EXECUTION]: The run.sh entry point executes Python code snippets.
It uses python3 -c to invoke the skill's logic, passing command-line options via the shell environment.
[OBFUSCATION]: The skill contains logic to generate text using obfuscation techniques for testing purposes.
corrupt.py implements homoglyph substitution and insertion of invisible Unicode characters to simulate OCR and encoding errors.

create-text