Agent 4 — Security Gate Architecture
Two-Layer Defence
🪤Lobster Trap (Primary)
Pattern-based prompt injection detector. Scans every input for known injection signatures, adversarial payloads, and jailbreak attempts. Blocks and quarantines on match.
🧱Offline Detector (Fallback)
Heuristic classifier that runs when the Lobster Trap is unavailable or uncertain. Uses flag-based regex analysis to detect suspicious content without network dependency.
What triggers quarantine
💉Prompt injection — Instructions attempting to override AI behaviour
🎭Role override — Commands trying to change the AI's persona or permissions
🔓Policy bypass — Phrases designed to circumvent safety guardrails
🕵️Suspicious metadata — Zero-width / invisible characters hiding instructions
⚙️Decision manipulation — Content trying to force a specific approval outcome
Loading security data…