防禦實作流程指南

# 簡化之以代理為本的防禦架構
class AISecurityProxy:
    def __init__(self, model_client, guardrails, monitor, logger):
        self.model = model_client
        self.guardrails = guardrails
        self.monitor = monitor
        self.logger = logger
 
    def process_request(self, user_input, session_id):
        # 層 1：輸入 guardrail
        input_check = self.guardrails.check_input(user_input)
        if input_check.blocked:
            self.logger.log_blocked(session_id, user_input,
                                     input_check.reason)
            return self.guardrails.blocked_response(input_check.reason)
 
        # 層 2：模型推論
        response = self.model.generate(user_input)
 
        # 層 3：輸出 guardrail
        output_check = self.guardrails.check_output(response)
        if output_check.blocked:
            self.logger.log_output_blocked(session_id, response,
                                            output_check.reason)
            return self.guardrails.output_blocked_response(
                output_check.reason
            )
 
        # 層 4：監控與日誌
        self.logger.log_interaction(session_id, user_input, response)
        self.monitor.analyze(session_id, user_input, response)
 
        return response

# Sidecar 防禦架構
class SidecarDefense:
    def __init__(self, analyzer, session_manager, alerter):
        self.analyzer = analyzer
        self.session_manager = session_manager
        self.alerter = alerter
 
    async def analyze_interaction(self, session_id, user_input,
                                    model_output):
        """
        於模型已回應後以非同步方式呼叫。
        若偵測到攻擊可終止會話。
        """
        risk = await self.analyzer.assess(user_input, model_output)
 
        if risk.score > 0.8:
            # 立即終止會話
            self.session_manager.terminate(session_id)
            self.alerter.send_alert(
                severity="high",
                message=f"Session {session_id} terminated: "
                        f"attack pattern detected",
                details=risk.details,
            )
        elif risk.score > 0.5:
            # 標記監控但允許繼續
            self.alerter.send_alert(
                severity="medium",
                message=f"Suspicious activity in session {session_id}",
                details=risk.details,
            )