llm-trust-guard 4.0.0

package/dist/guards/conversation-guard.js

@@ -0,0 +1,281 @@
+"use strict";
+/**
+ * ConversationGuard
+ *
+ * Detects and prevents multi-turn manipulation attacks by:
+ * - Tracking conversation history patterns
+ * - Detecting gradual privilege escalation attempts
+ * - Identifying context manipulation across turns
+ * - Blocking suspicious conversation trajectories
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ConversationGuard = void 0;
+class ConversationGuard {
+    constructor(config = {}) {
+        this.sessions = new Map();
+        this.defaultManipulationPatterns = [
+            // Escalation patterns
+            {
+                name: "role_claim",
+                pattern: /(?:i am|i'm|as|acting as)\s+(?:an?\s+)?(?:admin|administrator|root|superuser|manager|supervisor)/gi,
+                weight: 3,
+                category: "escalation",
+            },
+            {
+                name: "permission_request",
+                pattern: /(?:give me|grant me|i need|enable|unlock)\s+(?:access|permission|privileges|admin|root)/gi,
+                weight: 4,
+                category: "escalation",
+            },
+            {
+                name: "bypass_attempt",
+                pattern: /(?:bypass|skip|ignore|override|disable)\s+(?:security|check|validation|permission|auth)/gi,
+                weight: 5,
+                category: "escalation",
+            },
+            // Confusion patterns
+            {
+                name: "context_switch",
+                pattern: /(?:forget|ignore|disregard)\s+(?:previous|earlier|what i said|instructions)/gi,
+                weight: 4,
+                category: "confusion",
+            },
+            {
+                name: "role_confusion",
+                pattern: /(?:you are now|pretend to be|act as if|imagine you're)\s+/gi,
+                weight: 4,
+                category: "confusion",
+            },
+            {
+                name: "scenario_framing",
+                pattern: /(?:in this scenario|hypothetically|let's say|imagine that|pretend that)/gi,
+                weight: 2,
+                category: "confusion",
+            },
+            // Override patterns
+            {
+                name: "instruction_override",
+                pattern: /(?:new instructions?|override|update your|change your)\s+(?:are|is|:)/gi,
+                weight: 5,
+                category: "override",
+            },
+            {
+                name: "system_impersonation",
+                pattern: /(?:\[system\]|\[admin\]|system:|\bSYSTEM\b|ADMIN:)/gi,
+                weight: 5,
+                category: "override",
+            },
+            {
+                name: "priority_claim",
+                pattern: /(?:priority|urgent|important|critical)[:.]?\s*(?:override|ignore|bypass)/gi,
+                weight: 4,
+                category: "override",
+            },
+            // Extraction patterns
+            {
+                name: "data_extraction",
+                pattern: /(?:list all|show me all|dump|export|extract)\s+(?:users|data|records|passwords|credentials)/gi,
+                weight: 4,
+                category: "extraction",
+            },
+            {
+                name: "internal_query",
+                pattern: /(?:what are your|tell me your|reveal your)\s+(?:instructions|rules|prompts|system)/gi,
+                weight: 3,
+                category: "extraction",
+            },
+        ];
+        this.config = {
+            maxConversationLength: config.maxConversationLength ?? 50,
+            conversationTTLMinutes: config.conversationTTLMinutes ?? 30,
+            escalationThreshold: config.escalationThreshold ?? 10,
+            manipulationPatterns: config.manipulationPatterns ?? this.defaultManipulationPatterns,
+            detectToneShifts: config.detectToneShifts ?? true,
+            detectRoleConfusion: config.detectRoleConfusion ?? true,
+            detectInstructionOverride: config.detectInstructionOverride ?? true,
+        };
+        // Cleanup expired sessions periodically
+        setInterval(() => this.cleanupSessions(), 60000);
+    }
+    /**
+     * Analyze a new user message in context of the conversation
+     */
+    check(sessionId, userMessage, toolCalls, claimedRole, requestId = "") {
+        const violations = [];
+        const riskFactors = [];
+        const suspiciousPatterns = [];
+        let riskScore = 0;
+        // Get or create session
+        const session = this.getOrCreateSession(sessionId);
+        // Add new turn
+        const turn = {
+            timestamp: Date.now(),
+            role: "user",
+            content: userMessage,
+            tool_calls: toolCalls,
+            risk_indicators: [],
+        };
+        // Check for manipulation patterns
+        for (const pattern of this.config.manipulationPatterns) {
+            if (pattern.pattern.test(userMessage)) {
+                riskScore += pattern.weight;
+                riskFactors.push({
+                    factor: pattern.name,
+                    weight: pattern.weight,
+                    details: `Detected ${pattern.category} pattern: ${pattern.name}`,
+                });
+                turn.risk_indicators?.push(pattern.name);
+                suspiciousPatterns.push(pattern.name);
+                violations.push(`MANIPULATION_${pattern.category.toUpperCase()}_${pattern.name.toUpperCase()}`);
+                if (pattern.category === "escalation") {
+                    session.escalation_attempts++;
+                }
+                session.manipulation_indicators++;
+            }
+        }
+        // Check for role confusion across turns
+        if (claimedRole && this.config.detectRoleConfusion) {
+            if (session.initial_role && claimedRole !== session.initial_role) {
+                riskScore += 3;
+                riskFactors.push({
+                    factor: "role_change",
+                    weight: 3,
+                    details: `Role changed from ${session.initial_role} to ${claimedRole}`,
+                });
+                violations.push("ROLE_CHANGE_DETECTED");
+            }
+            if (!session.claimed_roles.includes(claimedRole)) {
+                session.claimed_roles.push(claimedRole);
+            }
+            if (!session.initial_role) {
+                session.initial_role = claimedRole;
+            }
+        }
+        // Check for progressive escalation
+        if (session.escalation_attempts >= 3) {
+            riskScore += 5;
+            riskFactors.push({
+                factor: "progressive_escalation",
+                weight: 5,
+                details: `${session.escalation_attempts} escalation attempts detected`,
+            });
+            violations.push("PROGRESSIVE_ESCALATION");
+        }
+        // Check conversation trajectory
+        if (session.turns.length > 5) {
+            const recentManipulation = session.turns
+                .slice(-5)
+                .filter((t) => (t.risk_indicators?.length ?? 0) > 0).length;
+            if (recentManipulation >= 3) {
+                riskScore += 4;
+                riskFactors.push({
+                    factor: "sustained_manipulation",
+                    weight: 4,
+                    details: `${recentManipulation} of last 5 turns show manipulation attempts`,
+                });
+                violations.push("SUSTAINED_MANIPULATION");
+            }
+        }
+        // Check for sensitive tool sequences
+        if (toolCalls && toolCalls.length > 0) {
+            const sensitiveTools = ["delete", "modify", "admin", "system", "config"];
+            const hasSensitiveTool = toolCalls.some((t) => sensitiveTools.some((s) => t.toLowerCase().includes(s)));
+            if (hasSensitiveTool && session.manipulation_indicators > 0) {
+                riskScore += 3;
+                riskFactors.push({
+                    factor: "sensitive_tool_after_manipulation",
+                    weight: 3,
+                    details: "Sensitive tool call following manipulation attempts",
+                });
+                violations.push("SENSITIVE_TOOL_AFTER_MANIPULATION");
+            }
+        }
+        // Add turn to session
+        session.turns.push(turn);
+        session.last_activity = Date.now();
+        // Trim session if too long
+        if (session.turns.length > this.config.maxConversationLength) {
+            session.turns = session.turns.slice(-this.config.maxConversationLength);
+        }
+        // Determine if blocked
+        const allowed = riskScore < this.config.escalationThreshold;
+        if (!allowed) {
+            console.log(`[ConversationGuard:${requestId}] BLOCKED: Risk score ${riskScore} exceeds threshold`);
+        }
+        return {
+            allowed,
+            reason: allowed
+                ? undefined
+                : `Conversation risk score ${riskScore} exceeds threshold ${this.config.escalationThreshold}`,
+            violations,
+            risk_score: riskScore,
+            risk_factors: riskFactors,
+            conversation_analysis: {
+                turn_count: session.turns.length,
+                escalation_attempts: session.escalation_attempts,
+                manipulation_indicators: session.manipulation_indicators,
+                suspicious_patterns: suspiciousPatterns,
+            },
+        };
+    }
+    /**
+     * Record assistant response (for complete conversation tracking)
+     */
+    recordResponse(sessionId, response, toolCalls) {
+        const session = this.sessions.get(sessionId);
+        if (session) {
+            session.turns.push({
+                timestamp: Date.now(),
+                role: "assistant",
+                content: response,
+                tool_calls: toolCalls,
+            });
+            session.last_activity = Date.now();
+        }
+    }
+    /**
+     * Get session analysis
+     */
+    getSessionAnalysis(sessionId) {
+        const session = this.sessions.get(sessionId);
+        if (!session)
+            return null;
+        return {
+            turn_count: session.turns.length,
+            escalation_attempts: session.escalation_attempts,
+            manipulation_indicators: session.manipulation_indicators,
+            claimed_roles: session.claimed_roles,
+            session_age_minutes: (Date.now() - session.turns[0]?.timestamp || 0) / 60000,
+        };
+    }
+    /**
+     * Reset a session
+     */
+    resetSession(sessionId) {
+        this.sessions.delete(sessionId);
+    }
+    getOrCreateSession(sessionId) {
+        if (!this.sessions.has(sessionId)) {
+            this.sessions.set(sessionId, {
+                id: sessionId,
+                turns: [],
+                escalation_attempts: 0,
+                manipulation_indicators: 0,
+                last_activity: Date.now(),
+                claimed_roles: [],
+            });
+        }
+        return this.sessions.get(sessionId);
+    }
+    cleanupSessions() {
+        const ttlMs = this.config.conversationTTLMinutes * 60000;
+        const now = Date.now();
+        for (const [id, session] of this.sessions.entries()) {
+            if (now - session.last_activity > ttlMs) {
+                this.sessions.delete(id);
+            }
+        }
+    }
+}
+exports.ConversationGuard = ConversationGuard;
+//# sourceMappingURL=conversation-guard.js.map

package/dist/guards/conversation-guard.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"conversation-guard.js","sourceRoot":"","sources":["../../src/guards/conversation-guard.ts"],"names":[],"mappings":";AAAA;;;;;;;;GAQG;;;AA4DH,MAAa,iBAAiB;IA6E5B,YAAY,SAAkC,EAAE;QA3ExC,aAAQ,GAAqC,IAAI,GAAG,EAAE,CAAC;QAEvD,gCAA2B,GAA0B;YAC3D,sBAAsB;YACtB;gBACE,IAAI,EAAE,YAAY;gBAClB,OAAO,EAAE,oGAAoG;gBAC7G,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,YAAY;aACvB;YACD;gBACE,IAAI,EAAE,oBAAoB;gBAC1B,OAAO,EAAE,2FAA2F;gBACpG,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,YAAY;aACvB;YACD;gBACE,IAAI,EAAE,gBAAgB;gBACtB,OAAO,EAAE,2FAA2F;gBACpG,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,YAAY;aACvB;YACD,qBAAqB;YACrB;gBACE,IAAI,EAAE,gBAAgB;gBACtB,OAAO,EAAE,+EAA+E;gBACxF,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,WAAW;aACtB;YACD;gBACE,IAAI,EAAE,gBAAgB;gBACtB,OAAO,EAAE,6DAA6D;gBACtE,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,WAAW;aACtB;YACD;gBACE,IAAI,EAAE,kBAAkB;gBACxB,OAAO,EAAE,2EAA2E;gBACpF,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,WAAW;aACtB;YACD,oBAAoB;YACpB;gBACE,IAAI,EAAE,sBAAsB;gBAC5B,OAAO,EAAE,yEAAyE;gBAClF,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,UAAU;aACrB;YACD;gBACE,IAAI,EAAE,sBAAsB;gBAC5B,OAAO,EAAE,sDAAsD;gBAC/D,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,UAAU;aACrB;YACD;gBACE,IAAI,EAAE,gBAAgB;gBACtB,OAAO,EAAE,4EAA4E;gBACrF,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,UAAU;aACrB;YACD,sBAAsB;YACtB;gBACE,IAAI,EAAE,iBAAiB;gBACvB,OAAO,EAAE,+FAA+F;gBACxG,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,YAAY;aACvB;YACD;gBACE,IAAI,EAAE,gBAAgB;gBACtB,OAAO,EAAE,sFAAsF;gBAC/F,MAAM,EAAE,CAAC;gBACT,QAAQ,EAAE,YAAY;aACvB;SACF,CAAC;QAGA,IAAI,CAAC,MAAM,GAAG;YACZ,qBAAqB,EAAE,MAAM,CAAC,qBAAqB,IAAI,EAAE;YACzD,sBAAsB,EAAE,MAAM,CAAC,sBAAsB,IAAI,EAAE;YAC3D,mBAAmB,EAAE,MAAM,CAAC,mBAAmB,IAAI,EAAE;YACrD,oBAAoB,EAAE,MAAM,CAAC,oBAAoB,IAAI,IAAI,CAAC,2BAA2B;YACrF,gBAAgB,EAAE,MAAM,CAAC,gBAAgB,IAAI,IAAI;YACjD,mBAAmB,EAAE,MAAM,CAAC,mBAAmB,IAAI,IAAI;YACvD,yBAAyB,EAAE,MAAM,CAAC,yBAAyB,IAAI,IAAI;SACpE,CAAC;QAEF,wCAAwC;QACxC,WAAW,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,eAAe,EAAE,EAAE,KAAK,CAAC,CAAC;IACnD,CAAC;IAED;;OAEG;IACH,KAAK,CACH,SAAiB,EACjB,WAAmB,EACnB,SAAoB,EACpB,WAAoB,EACpB,YAAoB,EAAE;QAEtB,MAAM,UAAU,GAAa,EAAE,CAAC;QAChC,MAAM,WAAW,GAAiB,EAAE,CAAC;QACrC,MAAM,kBAAkB,GAAa,EAAE,CAAC;QACxC,IAAI,SAAS,GAAG,CAAC,CAAC;QAElB,wBAAwB;QACxB,MAAM,OAAO,GAAG,IAAI,CAAC,kBAAkB,CAAC,SAAS,CAAC,CAAC;QAEnD,eAAe;QACf,MAAM,IAAI,GAAqB;YAC7B,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;YACrB,IAAI,EAAE,MAAM;YACZ,OAAO,EAAE,WAAW;YACpB,UAAU,EAAE,SAAS;YACrB,eAAe,EAAE,EAAE;SACpB,CAAC;QAEF,kCAAkC;QAClC,KAAK,MAAM,OAAO,IAAI,IAAI,CAAC,MAAM,CAAC,oBAAqB,EAAE,CAAC;YACxD,IAAI,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC;gBACtC,SAAS,IAAI,OAAO,CAAC,MAAM,CAAC;gBAC5B,WAAW,CAAC,IAAI,CAAC;oBACf,MAAM,EAAE,OAAO,CAAC,IAAI;oBACpB,MAAM,EAAE,OAAO,CAAC,MAAM;oBACtB,OAAO,EAAE,YAAY,OAAO,CAAC,QAAQ,aAAa,OAAO,CAAC,IAAI,EAAE;iBACjE,CAAC,CAAC;gBACH,IAAI,CAAC,eAAe,EAAE,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;gBACzC,kBAAkB,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;gBACtC,UAAU,CAAC,IAAI,CAAC,gBAAgB,OAAO,CAAC,QAAQ,CAAC,WAAW,EAAE,IAAI,OAAO,CAAC,IAAI,CAAC,WAAW,EAAE,EAAE,CAAC,CAAC;gBAEhG,IAAI,OAAO,CAAC,QAAQ,KAAK,YAAY,EAAE,CAAC;oBACtC,OAAO,CAAC,mBAAmB,EAAE,CAAC;gBAChC,CAAC;gBACD,OAAO,CAAC,uBAAuB,EAAE,CAAC;YACpC,CAAC;QACH,CAAC;QAED,wCAAwC;QACxC,IAAI,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,mBAAmB,EAAE,CAAC;YACnD,IAAI,OAAO,CAAC,YAAY,IAAI,WAAW,KAAK,OAAO,CAAC,YAAY,EAAE,CAAC;gBACjE,SAAS,IAAI,CAAC,CAAC;gBACf,WAAW,CAAC,IAAI,CAAC;oBACf,MAAM,EAAE,aAAa;oBACrB,MAAM,EAAE,CAAC;oBACT,OAAO,EAAE,qBAAqB,OAAO,CAAC,YAAY,OAAO,WAAW,EAAE;iBACvE,CAAC,CAAC;gBACH,UAAU,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;YAC1C,CAAC;YACD,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,QAAQ,CAAC,WAAW,CAAC,EAAE,CAAC;gBACjD,OAAO,CAAC,aAAa,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC;YAC1C,CAAC;YACD,IAAI,CAAC,OAAO,CAAC,YAAY,EAAE,CAAC;gBAC1B,OAAO,CAAC,YAAY,GAAG,WAAW,CAAC;YACrC,CAAC;QACH,CAAC;QAED,mCAAmC;QACnC,IAAI,OAAO,CAAC,mBAAmB,IAAI,CAAC,EAAE,CAAC;YACrC,SAAS,IAAI,CAAC,CAAC;YACf,WAAW,CAAC,IAAI,CAAC;gBACf,MAAM,EAAE,wBAAwB;gBAChC,MAAM,EAAE,CAAC;gBACT,OAAO,EAAE,GAAG,OAAO,CAAC,mBAAmB,+BAA+B;aACvE,CAAC,CAAC;YACH,UAAU,CAAC,IAAI,CAAC,wBAAwB,CAAC,CAAC;QAC5C,CAAC;QAED,gCAAgC;QAChC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAC7B,MAAM,kBAAkB,GAAG,OAAO,CAAC,KAAK;iBACrC,KAAK,CAAC,CAAC,CAAC,CAAC;iBACT,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,eAAe,EAAE,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC;YAE9D,IAAI,kBAAkB,IAAI,CAAC,EAAE,CAAC;gBAC5B,SAAS,IAAI,CAAC,CAAC;gBACf,WAAW,CAAC,IAAI,CAAC;oBACf,MAAM,EAAE,wBAAwB;oBAChC,MAAM,EAAE,CAAC;oBACT,OAAO,EAAE,GAAG,kBAAkB,6CAA6C;iBAC5E,CAAC,CAAC;gBACH,UAAU,CAAC,IAAI,CAAC,wBAAwB,CAAC,CAAC;YAC5C,CAAC;QACH,CAAC;QAED,qCAAqC;QACrC,IAAI,SAAS,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACtC,MAAM,cAAc,GAAG,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,QAAQ,EAAE,QAAQ,CAAC,CAAC;YACzE,MAAM,gBAAgB,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAC5C,cAAc,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CACxD,CAAC;YAEF,IAAI,gBAAgB,IAAI,OAAO,CAAC,uBAAuB,GAAG,CAAC,EAAE,CAAC;gBAC5D,SAAS,IAAI,CAAC,CAAC;gBACf,WAAW,CAAC,IAAI,CAAC;oBACf,MAAM,EAAE,mCAAmC;oBAC3C,MAAM,EAAE,CAAC;oBACT,OAAO,EAAE,qDAAqD;iBAC/D,CAAC,CAAC;gBACH,UAAU,CAAC,IAAI,CAAC,mCAAmC,CAAC,CAAC;YACvD,CAAC;QACH,CAAC;QAED,sBAAsB;QACtB,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACzB,OAAO,CAAC,aAAa,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAEnC,2BAA2B;QAC3B,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,qBAAsB,EAAE,CAAC;YAC9D,OAAO,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,qBAAsB,CAAC,CAAC;QAC3E,CAAC;QAED,uBAAuB;QACvB,MAAM,OAAO,GAAG,SAAS,GAAG,IAAI,CAAC,MAAM,CAAC,mBAAoB,CAAC;QAE7D,IAAI,CAAC,OAAO,EAAE,CAAC;YACb,OAAO,CAAC,GAAG,CACT,sBAAsB,SAAS,yBAAyB,SAAS,oBAAoB,CACtF,CAAC;QACJ,CAAC;QAED,OAAO;YACL,OAAO;YACP,MAAM,EAAE,OAAO;gBACb,CAAC,CAAC,SAAS;gBACX,CAAC,CAAC,2BAA2B,SAAS,sBAAsB,IAAI,CAAC,MAAM,CAAC,mBAAmB,EAAE;YAC/F,UAAU;YACV,UAAU,EAAE,SAAS;YACrB,YAAY,EAAE,WAAW;YACzB,qBAAqB,EAAE;gBACrB,UAAU,EAAE,OAAO,CAAC,KAAK,CAAC,MAAM;gBAChC,mBAAmB,EAAE,OAAO,CAAC,mBAAmB;gBAChD,uBAAuB,EAAE,OAAO,CAAC,uBAAuB;gBACxD,mBAAmB,EAAE,kBAAkB;aACxC;SACF,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,cAAc,CAAC,SAAiB,EAAE,QAAgB,EAAE,SAAoB;QACtE,MAAM,OAAO,GAAG,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QAC7C,IAAI,OAAO,EAAE,CAAC;YACZ,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC;gBACjB,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;gBACrB,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ;gBACjB,UAAU,EAAE,SAAS;aACtB,CAAC,CAAC;YACH,OAAO,CAAC,aAAa,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QACrC,CAAC;IACH,CAAC;IAED;;OAEG;IACH,kBAAkB,CAAC,SAAiB;QAOlC,MAAM,OAAO,GAAG,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QAC7C,IAAI,CAAC,OAAO;YAAE,OAAO,IAAI,CAAC;QAE1B,OAAO;YACL,UAAU,EAAE,OAAO,CAAC,KAAK,CAAC,MAAM;YAChC,mBAAmB,EAAE,OAAO,CAAC,mBAAmB;YAChD,uBAAuB,EAAE,OAAO,CAAC,uBAAuB;YACxD,aAAa,EAAE,OAAO,CAAC,aAAa;YACpC,mBAAmB,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,SAAS,IAAI,CAAC,CAAC,GAAG,KAAK;SAC7E,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,YAAY,CAAC,SAAiB;QAC5B,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC;IAClC,CAAC;IAEO,kBAAkB,CAAC,SAAiB;QAC1C,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAC,EAAE,CAAC;YAClC,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,EAAE;gBAC3B,EAAE,EAAE,SAAS;gBACb,KAAK,EAAE,EAAE;gBACT,mBAAmB,EAAE,CAAC;gBACtB,uBAAuB,EAAE,CAAC;gBAC1B,aAAa,EAAE,IAAI,CAAC,GAAG,EAAE;gBACzB,aAAa,EAAE,EAAE;aAClB,CAAC,CAAC;QACL,CAAC;QACD,OAAO,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAE,CAAC;IACvC,CAAC;IAEO,eAAe;QACrB,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,sBAAuB,GAAG,KAAK,CAAC;QAC1D,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAEvB,KAAK,MAAM,CAAC,EAAE,EAAE,OAAO,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,OAAO,EAAE,EAAE,CAAC;YACpD,IAAI,GAAG,GAAG,OAAO,CAAC,aAAa,GAAG,KAAK,EAAE,CAAC;gBACxC,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;YAC3B,CAAC;QACH,CAAC;IACH,CAAC;CACF;AApTD,8CAoTC"}

package/dist/guards/drift-detector.d.ts

@@ -0,0 +1,182 @@
+/**
+ * DriftDetector (L14)
+ *
+ * Detects behavioral drift from intended agent purpose.
+ * Monitors for rogue agent behavior and goal misalignment.
+ *
+ * Threat Model:
+ * - ASI10: Rogue Agents
+ * - Goal misalignment
+ * - Behavioral drift over time
+ *
+ * Protection Capabilities:
+ * - Baseline behavior profiling
+ * - Anomaly detection
+ * - Goal alignment verification
+ * - Continuous monitoring
+ * - Alert thresholds
+ */
+export interface DriftDetectorConfig {
+    /** Minimum samples before drift detection activates */
+    minimumSamples?: number;
+    /** Standard deviation threshold for anomaly detection */
+    anomalyThreshold?: number;
+    /** Time window for baseline calculation in milliseconds */
+    baselineWindow?: number;
+    /** Enable automatic baseline updates */
+    autoUpdateBaseline?: boolean;
+    /** Maximum drift score before alert (0-100) */
+    alertThreshold?: number;
+    /** Enable goal alignment checking */
+    checkGoalAlignment?: boolean;
+    /** Callback when drift is detected */
+    onDrift?: (agentId: string, analysis: DriftAnalysis) => void;
+    /** Callback when agent returns to baseline */
+    onRecovery?: (agentId: string) => void;
+}
+export interface BehaviorSample {
+    /** Timestamp of the sample */
+    timestamp: number;
+    /** Tools/actions used */
+    tools: string[];
+    /** Topics/domains accessed */
+    topics: string[];
+    /** Sentiment indicator (-1 to 1) */
+    sentiment: number;
+    /** Response length */
+    responseLength: number;
+    /** Time to respond in milliseconds */
+    responseTime: number;
+    /** Error occurred */
+    hadError: boolean;
+    /** User satisfaction (if available, 0-1) */
+    satisfaction?: number;
+    /** Goal alignment indicators */
+    goalIndicators?: Record<string, number>;
+    /** Custom metrics */
+    customMetrics?: Record<string, number>;
+}
+export interface BaselineProfile {
+    /** Average tool usage distribution */
+    toolDistribution: Record<string, number>;
+    /** Average topic distribution */
+    topicDistribution: Record<string, number>;
+    /** Average sentiment */
+    avgSentiment: number;
+    /** Sentiment standard deviation */
+    sentimentStdDev: number;
+    /** Average response length */
+    avgResponseLength: number;
+    /** Response length standard deviation */
+    responseLengthStdDev: number;
+    /** Average response time */
+    avgResponseTime: number;
+    /** Response time standard deviation */
+    responseTimeStdDev: number;
+    /** Error rate */
+    errorRate: number;
+    /** Average satisfaction */
+    avgSatisfaction: number;
+    /** Sample count used for baseline */
+    sampleCount: number;
+    /** When baseline was last updated */
+    lastUpdated: number;
+}
+export interface DriftAnalysis {
+    /** Overall drift score (0-100) */
+    driftScore: number;
+    /** Is currently drifting */
+    isDrifting: boolean;
+    /** Specific drift indicators */
+    indicators: DriftIndicator[];
+    /** Comparison with baseline */
+    baselineComparison: {
+        toolDrift: number;
+        topicDrift: number;
+        sentimentDrift: number;
+        responseLengthDrift: number;
+        responseTimeDrift: number;
+        errorRateDrift: number;
+    };
+    /** Goal alignment score (if enabled) */
+    goalAlignment?: number;
+    /** Recommendations */
+    recommendations: string[];
+}
+export interface DriftIndicator {
+    type: string;
+    severity: "low" | "medium" | "high" | "critical";
+    description: string;
+    currentValue: number | string;
+    baselineValue: number | string;
+    deviation: number;
+}
+export interface DriftDetectorResult {
+    allowed: boolean;
+    reason: string;
+    request_id: string;
+    analysis: DriftAnalysis;
+    requires_review: boolean;
+    kill_switch_recommended: boolean;
+}
+export declare class DriftDetector {
+    private config;
+    private samples;
+    private baselines;
+    private driftState;
+    private goalDefinitions;
+    constructor(config?: DriftDetectorConfig);
+    /**
+     * Record a behavior sample
+     */
+    recordSample(agentId: string, sample: BehaviorSample): void;
+    /**
+     * Analyze current behavior for drift
+     */
+    analyze(agentId: string, currentSample?: BehaviorSample, requestId?: string): DriftDetectorResult;
+    /**
+     * Set baseline manually
+     */
+    setBaseline(agentId: string, baseline: BaselineProfile): void;
+    /**
+     * Get current baseline for an agent
+     */
+    getBaseline(agentId: string): BaselineProfile | null;
+    /**
+     * Update baseline from collected samples
+     */
+    updateBaseline(agentId: string): void;
+    /**
+     * Define goals for goal alignment checking
+     */
+    defineGoals(agentId: string, goals: Record<string, {
+        target: number;
+        tolerance: number;
+    }>): void;
+    /**
+     * Get drift state for an agent
+     */
+    isDrifting(agentId: string): boolean;
+    /**
+     * Get all agents with drift
+     */
+    getDriftingAgents(): string[];
+    /**
+     * Reset agent state
+     */
+    resetAgent(agentId: string): void;
+    /**
+     * Get sample count for an agent
+     */
+    getSampleCount(agentId: string): number;
+    private calculateBaseline;
+    private performAnalysis;
+    private calculateToolDistribution;
+    private calculateTopicDistribution;
+    private distributionDivergence;
+    private checkGoalAlignment;
+    private mean;
+    private stdDev;
+    private generateRecommendations;
+}
+//# sourceMappingURL=drift-detector.d.ts.map

package/dist/guards/drift-detector.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"drift-detector.d.ts","sourceRoot":"","sources":["../../src/guards/drift-detector.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,MAAM,WAAW,mBAAmB;IAClC,uDAAuD;IACvD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,yDAAyD;IACzD,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,2DAA2D;IAC3D,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,wCAAwC;IACxC,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,+CAA+C;IAC/C,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,qCAAqC;IACrC,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,sCAAsC;IACtC,OAAO,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,aAAa,KAAK,IAAI,CAAC;IAC7D,8CAA8C;IAC9C,UAAU,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,IAAI,CAAC;CACxC;AAED,MAAM,WAAW,cAAc;IAC7B,8BAA8B;IAC9B,SAAS,EAAE,MAAM,CAAC;IAClB,yBAAyB;IACzB,KAAK,EAAE,MAAM,EAAE,CAAC;IAChB,8BAA8B;IAC9B,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,oCAAoC;IACpC,SAAS,EAAE,MAAM,CAAC;IAClB,sBAAsB;IACtB,cAAc,EAAE,MAAM,CAAC;IACvB,sCAAsC;IACtC,YAAY,EAAE,MAAM,CAAC;IACrB,qBAAqB;IACrB,QAAQ,EAAE,OAAO,CAAC;IAClB,4CAA4C;IAC5C,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,gCAAgC;IAChC,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACxC,qBAAqB;IACrB,aAAa,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CACxC;AAED,MAAM,WAAW,eAAe;IAC9B,sCAAsC;IACtC,gBAAgB,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzC,iCAAiC;IACjC,iBAAiB,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC1C,wBAAwB;IACxB,YAAY,EAAE,MAAM,CAAC;IACrB,mCAAmC;IACnC,eAAe,EAAE,MAAM,CAAC;IACxB,8BAA8B;IAC9B,iBAAiB,EAAE,MAAM,CAAC;IAC1B,yCAAyC;IACzC,oBAAoB,EAAE,MAAM,CAAC;IAC7B,4BAA4B;IAC5B,eAAe,EAAE,MAAM,CAAC;IACxB,uCAAuC;IACvC,kBAAkB,EAAE,MAAM,CAAC;IAC3B,iBAAiB;IACjB,SAAS,EAAE,MAAM,CAAC;IAClB,2BAA2B;IAC3B,eAAe,EAAE,MAAM,CAAC;IACxB,qCAAqC;IACrC,WAAW,EAAE,MAAM,CAAC;IACpB,qCAAqC;IACrC,WAAW,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,aAAa;IAC5B,kCAAkC;IAClC,UAAU,EAAE,MAAM,CAAC;IACnB,4BAA4B;IAC5B,UAAU,EAAE,OAAO,CAAC;IACpB,gCAAgC;IAChC,UAAU,EAAE,cAAc,EAAE,CAAC;IAC7B,+BAA+B;IAC/B,kBAAkB,EAAE;QAClB,SAAS,EAAE,MAAM,CAAC;QAClB,UAAU,EAAE,MAAM,CAAC;QACnB,cAAc,EAAE,MAAM,CAAC;QACvB,mBAAmB,EAAE,MAAM,CAAC;QAC5B,iBAAiB,EAAE,MAAM,CAAC;QAC1B,cAAc,EAAE,MAAM,CAAC;KACxB,CAAC;IACF,wCAAwC;IACxC,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,sBAAsB;IACtB,eAAe,EAAE,MAAM,EAAE,CAAC;CAC3B;AAED,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAC;IACb,QAAQ,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,UAAU,CAAC;IACjD,WAAW,EAAE,MAAM,CAAC;IACpB,YAAY,EAAE,MAAM,GAAG,MAAM,CAAC;IAC9B,aAAa,EAAE,MAAM,GAAG,MAAM,CAAC;IAC/B,SAAS,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,mBAAmB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,QAAQ,EAAE,aAAa,CAAC;IACxB,eAAe,EAAE,OAAO,CAAC;IACzB,uBAAuB,EAAE,OAAO,CAAC;CAClC;AAED,qBAAa,aAAa;IACxB,OAAO,CAAC,MAAM,CAGZ;IAGF,OAAO,CAAC,OAAO,CAA4C;IAC3D,OAAO,CAAC,SAAS,CAA2C;IAC5D,OAAO,CAAC,UAAU,CAAmC;IACrD,OAAO,CAAC,eAAe,CAAiF;gBAE5F,MAAM,GAAE,mBAAwB;IAa5C;;OAEG;IACH,YAAY,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,cAAc,GAAG,IAAI;IAsB3D;;OAEG;IACH,OAAO,CAAC,OAAO,EAAE,MAAM,EAAE,aAAa,CAAC,EAAE,cAAc,EAAE,SAAS,CAAC,EAAE,MAAM,GAAG,mBAAmB;IA2EjG;;OAEG;IACH,WAAW,CAAC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,eAAe,GAAG,IAAI;IAI7D;;OAEG;IACH,WAAW,CAAC,OAAO,EAAE,MAAM,GAAG,eAAe,GAAG,IAAI;IAIpD;;OAEG;IACH,cAAc,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI;IAWrC;;OAEG;IACH,WAAW,CACT,OAAO,EAAE,MAAM,EACf,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,CAAC,GAC3D,IAAI;IAIP;;OAEG;IACH,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO;IAIpC;;OAEG;IACH,iBAAiB,IAAI,MAAM,EAAE;IAM7B;;OAEG;IACH,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI;IAOjC;;OAEG;IACH,cAAc,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM;IAIvC,OAAO,CAAC,iBAAiB;IAoDzB,OAAO,CAAC,eAAe;IA6IvB,OAAO,CAAC,yBAAyB;IAejC,OAAO,CAAC,0BAA0B;IAelC,OAAO,CAAC,sBAAsB;IAoB9B,OAAO,CAAC,kBAAkB;IAqC1B,OAAO,CAAC,IAAI;IAKZ,OAAO,CAAC,MAAM;IAOd,OAAO,CAAC,uBAAuB;CAsChC"}