The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external documents (PDF, Word, Images) and returns the extracted text to the agent. This is an inherent surface for indirect prompt injection attacks where a document could contain malicious instructions.
Ingestion points: SKILL.md (via upload_and_poll and upload function calls).
Boundary markers: Absent. The instructions do not explicitly tell the agent to ignore instructions embedded within the document content.
Capability inventory: File system read access (open, fs.createReadStream, os.listdir) and network access to the vendor's API.
Sanitization: Absent. The skill passes extracted text directly to the agent context.
[EXTERNAL_DOWNLOADS]: The skill instructions include standard package installations for the vendor's official SDKs.
Python: pip install mixedbread installs the official Python client.
TypeScript: npm install @mixedbread/sdk installs the official Node.js SDK.
[DATA_EXFILTRATION]: The skill uses the vendor's API at mixedbread.com to process documents. This is a standard operation for a cloud-based parsing service and uses an environment variable MXBAI_API_KEY for authentication.

mixedbread-parsing