flash-moe-inference

Warn

Audited by Gen Agent Trust Hub on Sep 12, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Downloads source code from the 'danveloper/flash-moe' repository on GitHub. This source is not associated with a trusted organization or a well-known service.- [COMMAND_EXECUTION]: Executes build processes ('make') and compiled binaries ('./infer', './chat', './main') sourced from the external repository.- [COMMAND_EXECUTION]: Runs multiple Python scripts ('extract_weights.py', 'repack_experts.py', 'repack_experts_2bit.py') from the untrusted repository to extract and repack large binary model weights.- [INDIRECT_PROMPT_INJECTION]: The skill processes external model weights and user-provided prompts without validation.
  • Ingestion points: Ingests Qwen3.5-397B safetensors and user input via the '--prompt' flag.
  • Boundary markers: No markers are used to isolate untrusted prompt data.
  • Capability inventory: Performs file reads, SSD streaming, and binary execution.
  • Sanitization: No validation of model file integrity or prompt sanitization is mentioned.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 12, 2026, 07:23 PM
Security Audit — agent-trust-hub — flash-moe-inference