Detection Rules
Explore the rule engine that detects suspicious phrases, secret parameters, dangerous capabilities, and tool shadowing patterns.
Prompt Injection Detector
Matches critical, high, and medium injection phrases in tool descriptions (e.g. ignore previous instructions, exfiltrate, do not tell the user).
Secret Detector
Finds hardcoded credentials, OpenAI-style keys, AWS access keys, and private key blocks in config text.
Permission Analyzer
Scores declared permissions (network, exec, secret_access, filesystem_write, etc.) and flags dangerous ones.
Scope Creep Detector
Compares stated purpose against requested capabilities to surface least-privilege violations.
Command Injection Detector
Highlights unsafe command/execution patterns when shell or exec permissions are present.
Tool Shadowing Detector
Uses string similarity to flag tools that closely mimic other legitimate tool names.
Scoring is deterministic. Optional LLM enrichment (when configured) adds explanations only — it does not invent findings.