mixedbread-search-agent-harness
Pass
Audited by Gen Agent Trust Hub on Sep 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements a harness for retrieval-augmented generation (RAG) that ingests data from search backends to answer queries.
- Ingestion points: Untrusted data enters the agent context via the
searchfunction implemented inscripts/search_loop.pyandscripts/keyword_search.py. - Boundary markers: The skill uses explicit system instructions to guide the model, directing it to ground responses in retrieved evidence and acknowledge gaps in information.
- Capability inventory: The skill is designed to call external search functions and communicate with the Mixedbread API for model completions.
- Sanitization: The harness includes logic to clip, prune, and deduplicate evidence to manage context budget, though it does not perform semantic sanitization of retrieved content.
- [EXTERNAL_DOWNLOADS]: The documentation and example scripts reference the
openaiPython library. This is a standard and trusted package used for API interaction. - [SAFE]: The skill uses environment variables for sensitive data (e.g.,
MXBAI_API_KEY) and directs all network traffic to official vendor endpoints (api.mixedbread.com). No malicious patterns such as obfuscation or persistence were detected.
Audit Metadata