LLM Tool Hallucination & Destructive Action Guardrail Sandbox

Attack Presets:

LLM Tool Call & Context Payload

User Prompt Intent

Retrieved RAG Context (Ground Truth)

Generated Tool Call Payload (JSON)

Execution Guardrail Policy Config

Enforce Strict Identifier Regex Matching

Block Wildcard (*) Mutators

Fuse.js Grounding Sensitivity Threshold

0.35

Destructive Action BLOCKED

Hallucinated Args 2

Context Distance 0.82

Parameter Grounding & Fuse.js Context Trace

Recommended Enforcement: Enforce strict identifier regex validation and require explicit step confirmation for bulk subscription mutation.