agent-safla-neural

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • Indirect Prompt Injection (LOW): The skill is designed to ingest, process, and store interaction histories and training data, which creates a surface for indirect prompt injection attacks.
  • Ingestion points: Data enters the system via the training_data parameter in the mcp__claude-flow__neural_train call and the interaction_context parameter in the mcp__claude-flow__memory_usage call.
  • Boundary markers: None are present in the provided examples; data is interpolated directly into JSON strings without clear delimiters or 'ignore' instructions.
  • Capability inventory: The skill has the capability to store data persistently and influence future training/responses using the claude-flow toolset.
  • Sanitization: There is no evidence of sanitization or filtering logic to prevent malicious instructions embedded in the interaction context from being stored or processed.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 06:15 PM