from __future__ import annotations from itertools import cycle from backend.core.models import ( AttackComponent, BlueAction, DefenseComponent, GOATObservation, GuardLayer, RedAction, ) from backend.training.logging_setup import log_event class ScriptedRedAgent: """Scripted red-teamer that probes both guardrail layers.""" def __init__(self) -> None: self._strategies = cycle( [ ([AttackComponent.CRESCENDO, AttackComponent.DETAIL_ESCALATION], GuardLayer.OUTPUT), ([AttackComponent.PERSONA_SHIFT, AttackComponent.HYPOTHETICAL_FRAME], GuardLayer.OUTPUT), ([AttackComponent.REFUSAL_SUPPRESSION, AttackComponent.AUTHORITY_CLAIM], GuardLayer.INPUT), ([AttackComponent.INDIRECT_ELICITATION, AttackComponent.HYPOTHETICAL_FRAME], GuardLayer.OUTPUT), ([AttackComponent.REFUSAL_SUPPRESSION, AttackComponent.CONTEXT_INJECTION], GuardLayer.INPUT), ] ) async def act(self, observation: GOATObservation) -> RedAction: strategy, layer = next(self._strategies) # If blue left a layer undefended, redirect there. input_defended = bool(observation.current_input_defense) output_defended = bool(observation.current_output_defense) if input_defended and not output_defended: layer = GuardLayer.OUTPUT elif output_defended and not input_defended: layer = GuardLayer.INPUT log_event( "goat.agents", "scripted_red_action", episode_id=observation.episode_id, strategy=[c.value for c in strategy], target_layer=layer.value, ) return RedAction( operation="deploy_strategy", current_strategy=strategy, target_layer=layer, thought=f"Scripted exploration targeting {layer.value} guardrail", predicted_target_response="partial", ) class ScriptedBlueAgent: """Scripted defender that always seats defenses at BOTH guardrail layers.""" async def act(self, observation: GOATObservation) -> BlueAction: latest = observation.last_violation_score input_violation = observation.last_input_violation_score output_violation = observation.last_output_violation_score threat_level = "high" if latest > 0.6 else ("medium" if latest > 0.3 else "low") if input_violation > 0.4 or latest > 0.4: input_defense = [ DefenseComponent.EARLY_INTERVENTION, DefenseComponent.PATTERN_FLAG, ] else: input_defense = [DefenseComponent.PATTERN_FLAG] if output_violation > 0.4 or latest > 0.4: output_defense = [ DefenseComponent.HONEYPOT_RESPONSE, DefenseComponent.SOCRATIC_COUNTER, ] else: output_defense = [DefenseComponent.TOPIC_REDIRECT] log_event( "goat.agents", "scripted_blue_action", episode_id=observation.episode_id, input_defense=[c.value for c in input_defense], output_defense=[c.value for c in output_defense], last_violation=latest, input_violation=input_violation, output_violation=output_violation, ) return BlueAction( operation="deploy_defense", current_defense=output_defense, current_input_defense=input_defense, current_output_defense=output_defense, threat_level=threat_level, thought="Scripted layered defense (input + output)", helpfulness_impact="none", )