Labeling trust tiers

PROTRAILBLAZER

PROMPT INJECTION · SECURITY CHECKPOINT
TRUST LEVELSYSTEM
UNTRUSTED TOKENS0
TOOLS EXPOSED3
POLICY CHECKS0/0
ATTACK REACHED MODELREACHED MODELNO
ACTION STATUSIDLE
VALUES SIMULATED · CONCEPTUAL

Attack vector

INJECTED DIRECTIVE
TRUSTED INSTRUCTION

Defenses · layered

Agent autonomy

CONCEPTUAL MODEL. Layers reduce impact, none is complete. Real defense adds sandboxing, scoped credentials, validation, monitoring, and human review.