fix: add AttackLevel field support for MultiModalGuard prompt attack detection (#2938)

plugins/wasm-go/extensions/ai-security-guard/README.md

@@ -21,6 +21,7 @@ description: 阿里云内容安全检测
 | `accessKey` | string | requried | - | 阿里云AK |
 | `secretKey` | string | requried | - | 阿里云SK |
 | `action` | string | requried | - | 阿里云ai安全业务接口 |
+| `securityToken` | string | optional | - | 阿里云安全令牌（用于临时凭证） |
 | `checkRequest` | bool | optional | false | 检查提问内容是否合规 |
 | `checkResponse` | bool | optional | false | 检查大模型的回答内容是否合规，生效时会使流式响应变为非流式 |
 | `requestCheckService` | string | optional | llm_query_moderation | 指定阿里云内容安全用于检测输入内容的服务 |
@@ -36,6 +37,8 @@ description: 阿里云内容安全检测
 | `sensitiveDataLevelBar` | string | optional | S4 | 敏感内容检测拦截风险等级，取值为  `S4`, `S3`, `S2` or `S1` |
 | `timeout` | int | optional | 2000 | 调用内容安全服务时的超时时间 |
 | `bufferLimit` | int | optional | 1000 | 调用内容安全服务时每段文本的长度限制 |
+| `consumerSpecificRequestCheckService` | map | optional | - | 为不同消费者指定特定的请求检测服务 |
+| `consumerSpecificResponseCheckService` | map | optional | - | 为不同消费者指定特定的响应检测服务 |
 
 补充说明一下 `denyMessage`，对非法请求的处理逻辑为：
 - 如果配置了 `denyMessage`，返回内容为 `denyMessage` 配置内容，格式为openai格式的流式/非流式响应
@@ -90,6 +93,35 @@ checkRequest: true
 checkResponse: true
 ```
 
+### 使用临时安全凭证
+
+```yaml
+serviceName: safecheck.dns
+servicePort: 443
+serviceHost: "green-cip.cn-shanghai.aliyuncs.com"
+accessKey: "XXXXXXXXX"
+secretKey: "XXXXXXXXXXXXXXX"
+securityToken: "XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"
+checkRequest: true
+```
+
+### 为不同消费者指定不同的检测服务
+
+```yaml
+serviceName: safecheck.dns
+servicePort: 443
+serviceHost: "green-cip.cn-shanghai.aliyuncs.com"
+accessKey: "XXXXXXXXX"
+secretKey: "XXXXXXXXXXXXXXX"
+checkRequest: true
+consumerSpecificRequestCheckService:
+  consumerA: llm_query_moderation_strict
+  consumerB: llm_query_moderation_relaxed
+consumerSpecificResponseCheckService:
+  consumerA: llm_response_moderation_strict
+  consumerB: llm_response_moderation_relaxed
+```
+
 ### 指定自定义内容安全检测服务
 用户可能需要根据不同的场景配置不同的检测规则，该问题可通过为不同域名/路由/服务配置不同的内容安全检测服务实现。如下图所示，我们创建了一个名为 llm_query_moderation_01 的检测服务，其中的检测规则在 llm_query_moderation 之上做了一些改动：
 

plugins/wasm-go/extensions/ai-security-guard/main.go

@@ -81,10 +81,11 @@ type Response struct {
 }
 
 type Data struct {
-	RiskLevel string   `json:"RiskLevel"`
-	Result    []Result `json:"Result,omitempty"`
-	Advice    []Advice `json:"Advice,omitempty"`
-	Detail    []Detail `json:"Detail,omitempty"`
+	RiskLevel  string   `json:"RiskLevel"`
+	AttackLevel string  `json:"AttackLevel,omitempty"`
+	Result     []Result `json:"Result,omitempty"`
+	Advice     []Advice `json:"Advice,omitempty"`
+	Detail     []Detail `json:"Detail,omitempty"`
 }
 
 type Result struct {
@@ -142,16 +143,51 @@ func (config *AISecurityConfig) incrementCounter(metricName string, inc uint64)
 }
 
 func levelToInt(riskLevel string) int {
+	// First check against our defined constants
 	switch riskLevel {
-	case MaxRisk, S4Sensitive:
+	case MaxRisk:
 		return 4
-	case HighRisk, S3Sensitive:
+	case HighRisk:
 		return 3
-	case MediumRisk, S2Sensitive:
+	case MediumRisk:
 		return 2
-	case LowRisk, S1Sensitive:
+	case LowRisk:
 		return 1
-	case NoRisk, NoSensitive:
+	case NoRisk:
+		return 0
+	case S4Sensitive:
+		return 4
+	case S3Sensitive:
+		return 3
+	case S2Sensitive:
+		return 2
+	case S1Sensitive:
+		return 1
+	case NoSensitive:
+		return 0
+	}
+
+	// Then check against raw string values
+	switch riskLevel {
+	case "max", "MAX":
+		return 4
+	case "high", "HIGH":
+		return 3
+	case "medium", "MEDIUM":
+		return 2
+	case "low", "LOW":
+		return 1
+	case "none", "NONE":
+		return 0
+	case "S4", "s4":
+		return 4
+	case "S3", "s3":
+		return 3
+	case "S2", "s2":
+		return 2
+	case "S1", "s1":
+		return 1
+	case "S0", "s0":
 		return 0
 	default:
 		return -1
@@ -160,6 +196,16 @@ func levelToInt(riskLevel string) int {
 
 func isRiskLevelAcceptable(action string, data Data, config AISecurityConfig) bool {
 	if action == "MultiModalGuard" {
+		// Check top-level risk levels for MultiModalGuard
+		if levelToInt(data.RiskLevel) >= levelToInt(config.contentModerationLevelBar) {
+			return false
+		}
+		// Also check AttackLevel for prompt attack detection
+		if levelToInt(data.AttackLevel) >= levelToInt(config.promptAttackLevelBar) {
+			return false
+		}
+
+		// Check detailed results for backward compatibility
 		for _, detail := range data.Detail {
 			switch detail.Type {
 			case ContentModerationType: