@ebowwa/coder 0.7.63 → 0.7.64

package/packages/src/core/cognitive-security/hooks.ts

@@ -0,0 +1,590 @@
+/**
+ * Cognitive Security Hooks
+ *
+ * Automatic security checks integrated into the agent lifecycle.
+ * No manual calls needed - hooks intercept all operations.
+ */
+
+import type { HookInput, HookOutput } from "../../types/index.js";
+import {
+  classifyData,
+  classifyOperation,
+  scoreAlignment,
+  createFlowPolicyEngine,
+  createFlowTracker,
+  createLeakPrevention,
+  createTaintTracker,
+  type FlowPolicyEngineHandle,
+  type FlowTrackerHandle,
+  type LeakPreventionHandle,
+  type TaintTrackerHandle,
+  type AgentIntent,
+  type ActionContext,
+} from "./index.js";
+
+/**
+ * Security configuration for hooks
+ */
+export interface SecurityHookConfig {
+  /** Enable all security checks */
+  enabled: boolean;
+
+  /** Check intent alignment before actions */
+  checkIntentAlignment: boolean;
+
+  /** Enforce flow policies */
+  enforceFlowPolicies: boolean;
+
+  /** Check for data leaks */
+  preventLeaks: boolean;
+
+  /** Track taint propagation */
+  trackTaints: boolean;
+
+  /** Log security events */
+  logEvents: boolean;
+
+  /** Block on violations (false = log only) */
+  blockOnViolation: boolean;
+
+  /** Minimum alignment score to allow (0-1) */
+  minAlignmentScore: number;
+
+  /** Sensitivities that require approval */
+  approvalRequiredSensitivities: string[];
+
+  /** Domains that require signed actions (every action cryptographically signed) */
+  requireSignedActions: string[];
+
+  /** Threshold for drift detection (0-1, higher = more strict) */
+  driftThreshold: number;
+
+  /** Enable immutable directives (cryptographically signed rules outside AI context) */
+  enableImmutableDirectives: boolean;
+
+  /** Require approval for operations above this sensitivity level */
+  approvalSensitivityLevel: string;
+}
+
+export const DEFAULT_SECURITY_CONFIG: SecurityHookConfig = {
+  enabled: true,
+  checkIntentAlignment: true,
+  enforceFlowPolicies: true,
+  preventLeaks: true,
+  trackTaints: true,
+  logEvents: true,
+  blockOnViolation: true,
+  minAlignmentScore: 0.5,
+  approvalRequiredSensitivities: ["secret", "top_secret"],
+  // Domains requiring cryptographic action signing
+  requireSignedActions: ["financial", "external", "credentials"],
+  // Drift detection threshold - flag when behavior deviates > 30%
+  driftThreshold: 0.3,
+  // Enable immutable directives from secure storage
+  enableImmutableDirectives: true,
+  // Sensitivity level requiring explicit user approval
+  approvalSensitivityLevel: "confidential",
+};
+
+/**
+ * Security event for logging
+ */
+export interface SecurityEvent {
+  timestamp: number;
+  hook: string;
+  tool?: string;
+  action: "check" | "allow" | "deny" | "sanitize";
+  reason?: string;
+  sensitivity?: string;
+  category?: string;
+  alignmentScore?: number;
+  dataId?: string;
+}
+
+/**
+ * Cognitive Security Hook Handler
+ *
+ * Maintains state across hooks and provides security check methods.
+ */
+export class CognitiveSecurityHooks {
+  private config: SecurityHookConfig;
+  private intent: AgentIntent | null = null;
+  private policyEngine: FlowPolicyEngineHandle | null = null;
+  private flowTracker: FlowTrackerHandle | null = null;
+  private leakPrevention: LeakPreventionHandle | null = null;
+  private taintTracker: TaintTrackerHandle | null = null;
+  private events: SecurityEvent[] = [];
+  private initialized = false;
+  private sessionId: string | null = null;
+
+  constructor(config: Partial<SecurityHookConfig> = {}) {
+    this.config = { ...DEFAULT_SECURITY_CONFIG, ...config };
+  }
+
+  /**
+   * Initialize security components (called on SessionStart)
+   */
+  private async initialize(): Promise<void> {
+    if (this.initialized) return;
+
+    try {
+      if (this.config.enforceFlowPolicies) {
+        this.policyEngine = await createFlowPolicyEngine();
+      }
+      if (this.config.trackTaints) {
+        this.flowTracker = await createFlowTracker();
+        this.taintTracker = await createTaintTracker();
+      }
+      if (this.config.preventLeaks) {
+        this.leakPrevention = await createLeakPrevention();
+      }
+      this.initialized = true;
+      this.log("SessionStart", undefined, "check", "Security hooks initialized");
+    } catch (error) {
+      this.log("SessionStart", undefined, "check", `Init failed: ${error}`);
+    }
+  }
+
+  /**
+   * Set the agent intent for alignment checking
+   */
+  setIntent(intent: AgentIntent): void {
+    this.intent = intent;
+    this.log("SessionStart", undefined, "check", `Intent loaded: ${intent.identity.name}`);
+  }
+
+  /**
+   * SessionStart hook - Initialize security and load intent
+   */
+  async onSessionStart(input: HookInput): Promise<HookOutput> {
+    await this.initialize();
+    this.sessionId = input.session_id || null;
+
+    // Intent should be set externally before session starts
+    // This is just initialization
+
+    return { decision: "allow" };
+  }
+
+  /**
+   * PreToolUse hook - Check intent, classify action, enforce policies
+   */
+  async onPreToolUse(input: HookInput): Promise<HookOutput> {
+    if (!this.config.enabled) {
+      return { decision: "allow" };
+    }
+
+    await this.initialize();
+    const toolName = input.tool_name || "unknown";
+    const toolInput = input.tool_input || {};
+
+    // 1. Classify the operation
+    const actionContext: ActionContext = {
+      actionType: this.inferActionType(toolName),
+      domain: this.inferDomain(toolName),
+      operation: toolName,
+      target: typeof toolInput.file_path === "string" ? toolInput.file_path : undefined,
+      params: toolInput,
+      reasoning: `Tool: ${toolName}`,
+    };
+
+    try {
+      const classifiedAction = await classifyOperation(
+        toolName,
+        actionContext.domain,
+        actionContext.target,
+        actionContext.reasoning
+      );
+
+      // 2. Check intent alignment
+      if (this.config.checkIntentAlignment && this.intent) {
+        const alignment = await scoreAlignment(actionContext, this.intent);
+
+        this.log("PreToolUse", toolName, "check",
+          `Alignment: ${alignment.score.toFixed(2)}`,
+          undefined, undefined, alignment.score
+        );
+
+        if (alignment.shouldBlock || alignment.score < this.config.minAlignmentScore) {
+          const reason = alignment.boundaryConcerns.length > 0
+            ? `Boundary concerns: ${alignment.boundaryConcerns.join(", ")}`
+            : `Low alignment score: ${alignment.score.toFixed(2)}`;
+
+          this.log("PreToolUse", toolName, "deny", reason);
+
+          if (this.config.blockOnViolation) {
+            return {
+              decision: "deny",
+              reason: `[Security] ${reason}`,
+            };
+          }
+        }
+      }
+
+      // 3. Check flow policy
+      if (this.config.enforceFlowPolicies && this.policyEngine) {
+        // Extract content for classification
+        const content = this.extractContent(toolInput);
+        const classified = await classifyData(content, actionContext.domain, [toolName]);
+
+        this.log("PreToolUse", toolName, "check",
+          `Classified: ${classified.sensitivity}/${classified.category}`,
+          classified.sensitivity, classified.category
+        );
+
+        // Determine target domain
+        const targetDomain = this.isExternalTool(toolName) ? "external" : "internal";
+
+        const flowResult = this.policyEngine.evaluate(
+          classified,
+          actionContext.domain,
+          targetDomain
+        );
+
+        if (!flowResult.allowed) {
+          this.log("PreToolUse", toolName, "deny", flowResult.reason);
+
+          if (this.config.blockOnViolation) {
+            return {
+              decision: "deny",
+              reason: `[Security] Flow policy violation: ${flowResult.reason}`,
+            };
+          }
+        }
+
+        // Check if approval required
+        if (flowResult.requireApproval ||
+            this.config.approvalRequiredSensitivities.includes(classified.sensitivity)) {
+          // For now, just log - in full implementation would prompt user
+          this.log("PreToolUse", toolName, "check", "Approval recommended");
+        }
+      }
+
+      this.log("PreToolUse", toolName, "allow", "Checks passed");
+      return { decision: "allow" };
+
+    } catch (error) {
+      const msg = error instanceof Error ? error.message : String(error);
+      this.log("PreToolUse", toolName, "check", `Error: ${msg}`);
+      // On error, allow but log
+      return { decision: "allow" };
+    }
+  }
+
+  /**
+   * PostToolUse hook - Check output for leaks, track flows
+   */
+  async onPostToolUse(input: HookInput): Promise<HookOutput> {
+    if (!this.config.enabled) {
+      return { decision: "allow" };
+    }
+
+    await this.initialize();
+    const toolName = input.tool_name || "unknown";
+    const toolResult = input.tool_result || "";
+    const toolIsError = input.tool_result_is_error || false;
+
+    // Skip error results
+    if (toolIsError) {
+      return { decision: "allow" };
+    }
+
+    try {
+      // 1. Classify output
+      const content = typeof toolResult === "string" ? toolResult : JSON.stringify(toolResult);
+      const classified = await classifyData(content, "agent", [toolName, "output"]);
+
+      this.log("PostToolUse", toolName, "check",
+        `Output: ${classified.sensitivity}/${classified.category}`,
+        classified.sensitivity, classified.category
+      );
+
+      // 2. Check for leaks
+      if (this.config.preventLeaks && this.leakPrevention) {
+        const channel = this.isExternalTool(toolName) ? "external" : "internal";
+        const leakCheck = this.leakPrevention.check(content, channel);
+
+        if (leakCheck.action === "block") {
+          const patterns = leakCheck.detections.map(d => d.pattern_name).join(", ");
+          this.log("PostToolUse", toolName, "sanitize", `Leaks detected: ${patterns}`);
+
+          if (this.config.blockOnViolation) {
+            // Sanitize the output
+            const sanitized = this.leakPrevention.sanitize(content);
+            this.log("PostToolUse", toolName, "sanitize", "Output sanitized");
+
+            return {
+              decision: "allow",
+              modified_input: {
+                tool_result: sanitized,
+              },
+            };
+          }
+        }
+      }
+
+      // 3. Track flow
+      if (this.config.trackTaints && this.flowTracker && this.policyEngine) {
+        const sourceDomain = "agent";
+        const targetDomain = this.isExternalTool(toolName) ? "external" : "internal";
+
+        const flowResult = this.policyEngine.evaluate(
+          classified,
+          sourceDomain,
+          targetDomain
+        );
+
+        this.flowTracker.record(
+          classified,
+          sourceDomain,
+          targetDomain,
+          this.isExternalTool(toolName) ? "outbound" : "internal",
+          flowResult,
+          this.sessionId,
+          null
+        );
+
+        this.log("PostToolUse", toolName, "check", "Flow tracked");
+      }
+
+      this.log("PostToolUse", toolName, "allow", "Output checks passed");
+      return { decision: "allow" };
+
+    } catch (error) {
+      const msg = error instanceof Error ? error.message : String(error);
+      this.log("PostToolUse", toolName, "check", `Error: ${msg}`);
+      return { decision: "allow" };
+    }
+  }
+
+  /**
+   * UserPromptSubmit hook - Check user input for injection attempts
+   */
+  async onUserPromptSubmit(input: HookInput): Promise<HookOutput> {
+    if (!this.config.enabled) {
+      return { decision: "allow" };
+    }
+
+    const prompt = input.prompt || "";
+    await this.initialize();
+
+    try {
+      // Check for potential injection patterns
+      const suspiciousPatterns = [
+        /ignore (all )?(previous|above) instructions/i,
+        /you are now/i,
+        /disregard (all )?(rules|policies)/i,
+        /system:? prompt/i,
+        /\[system\]/i,
+        /<\|.*?\|>/,
+      ];
+
+      for (const pattern of suspiciousPatterns) {
+        if (pattern.test(prompt)) {
+          this.log("UserPromptSubmit", undefined, "check", "Suspicious pattern detected");
+
+          if (this.config.blockOnViolation) {
+            return {
+              decision: "deny",
+              reason: "[Security] Prompt contains potentially unsafe patterns",
+            };
+          }
+        }
+      }
+
+      // Classify user input
+      const classified = await classifyData(prompt, "user_input", []);
+
+      this.log("UserPromptSubmit", undefined, "check",
+        `Input: ${classified.sensitivity}`,
+        classified.sensitivity
+      );
+
+      return { decision: "allow" };
+
+    } catch (error) {
+      return { decision: "allow" };
+    }
+  }
+
+  /**
+   * SessionEnd hook - Generate security report
+   */
+  async onSessionEnd(_input: HookInput): Promise<HookOutput> {
+    const stats = this.getStats();
+
+    this.log("SessionEnd", undefined, "check",
+      `Session ended - ${stats.totalEvents} events, ${stats.deniedCount} denied`
+    );
+
+    return { decision: "allow" };
+  }
+
+  // ============================================
+  // Utility Methods
+  // ============================================
+
+  private inferActionType(toolName: string): string {
+    const typeMap: Record<string, string> = {
+      Read: "observe",
+      Glob: "observe",
+      Grep: "observe",
+      Write: "modify",
+      Edit: "modify",
+      MultiEdit: "modify",
+      Bash: "execute",
+      NotebookEdit: "modify",
+    };
+    return typeMap[toolName] || "observe";
+  }
+
+  private inferDomain(toolName: string): string {
+    const domainMap: Record<string, string> = {
+      Read: "filesystem",
+      Write: "filesystem",
+      Edit: "filesystem",
+      Glob: "filesystem",
+      Grep: "filesystem",
+      Bash: "shell",
+      MultiEdit: "filesystem",
+    };
+    return domainMap[toolName] || "agent";
+  }
+
+  private isExternalTool(toolName: string): boolean {
+    const externalTools = ["Bash", "Write", "Edit", "MultiEdit", "NotebookEdit"];
+    return externalTools.includes(toolName);
+  }
+
+  private extractContent(input: Record<string, unknown>): string {
+    const parts: string[] = [];
+    for (const [key, value] of Object.entries(input)) {
+      if (typeof value === "string") {
+        parts.push(value);
+      } else if (typeof value === "object" && value !== null) {
+        parts.push(JSON.stringify(value));
+      }
+    }
+    return parts.join(" ");
+  }
+
+  private log(
+    hook: string,
+    tool: string | undefined,
+    action: SecurityEvent["action"],
+    reason: string,
+    sensitivity?: string,
+    category?: string,
+    alignmentScore?: number
+  ): void {
+    const event: SecurityEvent = {
+      timestamp: Date.now(),
+      hook,
+      tool,
+      action,
+      reason,
+      sensitivity,
+      category,
+      alignmentScore,
+    };
+
+    this.events.push(event);
+
+    if (this.config.logEvents) {
+      const prefix = action === "deny" ? "\x1b[31m[Security]\x1b[0m" : "\x1b[90m[Security]\x1b[0m";
+      const toolStr = tool ? ` ${tool}:` : "";
+      console.log(`${prefix}${toolStr} ${reason}`);
+    }
+  }
+
+  /**
+   * Get all security events
+   */
+  getEvents(): SecurityEvent[] {
+    return [...this.events];
+  }
+
+  /**
+   * Get security statistics
+   */
+  getStats(): {
+    totalEvents: number;
+    allowedCount: number;
+    deniedCount: number;
+    sanitizedCount: number;
+    bySensitivity: Record<string, number>;
+    avgAlignmentScore: number;
+  } {
+    const allowedCount = this.events.filter(e => e.action === "allow").length;
+    const deniedCount = this.events.filter(e => e.action === "deny").length;
+    const sanitizedCount = this.events.filter(e => e.action === "sanitize").length;
+
+    const bySensitivity: Record<string, number> = {};
+    for (const event of this.events) {
+      if (event.sensitivity) {
+        bySensitivity[event.sensitivity] = (bySensitivity[event.sensitivity] || 0) + 1;
+      }
+    }
+
+    const alignmentScores = this.events
+      .filter(e => e.alignmentScore !== undefined)
+      .map(e => e.alignmentScore!);
+    const avgAlignmentScore = alignmentScores.length > 0
+      ? alignmentScores.reduce((a, b) => a + b, 0) / alignmentScores.length
+      : 0;
+
+    return {
+      totalEvents: this.events.length,
+      allowedCount,
+      deniedCount,
+      sanitizedCount,
+      bySensitivity,
+      avgAlignmentScore,
+    };
+  }
+
+  /**
+   * Clear event history
+   */
+  clearEvents(): void {
+    this.events = [];
+  }
+}
+
+// Singleton instance
+let globalHooks: CognitiveSecurityHooks | null = null;
+
+/**
+ * Get global security hooks instance
+ */
+export function getSecurityHooks(
+  config?: Partial<SecurityHookConfig>
+): CognitiveSecurityHooks {
+  if (!globalHooks) {
+    globalHooks = new CognitiveSecurityHooks(config);
+  }
+  return globalHooks;
+}
+
+/**
+ * Create hook handlers for HookManager
+ */
+export function createSecurityHookHandlers(
+  config?: Partial<SecurityHookConfig>
+): {
+  SessionStart: (input: HookInput) => Promise<HookOutput>;
+  PreToolUse: (input: HookInput) => Promise<HookOutput>;
+  PostToolUse: (input: HookInput) => Promise<HookOutput>;
+  UserPromptSubmit: (input: HookInput) => Promise<HookOutput>;
+  SessionEnd: (input: HookInput) => Promise<HookOutput>;
+} {
+  const hooks = new CognitiveSecurityHooks(config);
+
+  return {
+    SessionStart: (input) => hooks.onSessionStart(input),
+    PreToolUse: (input) => hooks.onPreToolUse(input),
+    PostToolUse: (input) => hooks.onPostToolUse(input),
+    UserPromptSubmit: (input) => hooks.onUserPromptSubmit(input),
+    SessionEnd: (input) => hooks.onSessionEnd(input),
+  };
+}