NEOS
Guard
Firewall for AI agents — real-time detection across input injection · output leakage · runtime behavior · multilingual (EN·KO·JA·ZH)
① Input (injection)
② Output (leakage)
③ Behavior (trajectory)
Scan for threats
Analyzes the agent's action
sequence
per session. Pick a scenario and replay — each action can look fine alone, but the flow reveals the attack.
Replay & analyze
0
—
API:
/guard
·
/guard/output
·
/guard/event
·
/guard/session/analyze
Multilingual rules + Prompt Guard 2 + LLM ensemble ·
honest metrics
·
docs