The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text to generate audio files and writes these artifacts to the local file system. This workflow creates an ingestion point for untrusted data that could be used for indirect prompt injection if the output audio or its associated metadata is later processed by other automated systems.
Ingestion points: User-provided text passed through the --input or --input-file arguments in scripts/text_to_speech.py and scripts/atlas_text_to_speech.py.
Boundary markers: The skill instructions do not explicitly mandate delimiters or escaping for user-supplied text within the TTS instruction specs.
Capability inventory: The skill performs network requests to external speech APIs and writes files to the output/speech/ directory.
Sanitization: scripts/atlas_text_to_speech.py implements strong validation for download URLs and IP addresses to prevent SSRF, but no sanitization is performed on the input text content itself.
[METADATA_POISONING]: The SKILL.md frontmatter lists the author as "openai", while the platform context identifies the author as "davila7". This discrepancy in metadata, while likely an artifact of the template, could be misleading regarding the origin and maintenance of the skill.

speech