still-qa
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted text from the
--promptargument and the--biblefile, which are interpolated directly into the system prompt for the VLM judge. \n - Ingestion points:
scripts/judge.pyreads the--promptCLI argument and the file content from the path in--bible. \n - Boundary markers: Absent. The data is concatenated with simple text labels without delimiters. \n
- Capability inventory:
scripts/judge.pyperforms file reads (open) and network operations (urllib.request). \n - Sanitization: Absent. No escaping or filtering is applied to the inputs. \n- [DATA_EXFILTRATION]: The script
scripts/judge.pyreads local file data and transmits it to a network endpoint. \n - It reads images and text files specified by CLI arguments. \n
- It sends this content, plus the
VLM_API_KEYenvironment variable, to the URL specified by--endpoint. \n - This allows for the exfiltration of local assets to non-whitelisted domains if the endpoint is changed from its default.
Audit Metadata