feat(social): moderation hooks & governance integration (BE-093)

.gitignore

@@ -96,3 +96,6 @@ commit_*.ps1
 commit_*.sh
 
 main
+branch_structure.json
+temp_auto_push.bat
+temp_interactive_push.bat

BackendAcademy/src/ai/prompt-guard.service.ts

@@ -0,0 +1,102 @@
+import { Injectable } from '@nestjs/common';
+
+/**
+ * Prompt-injection & content-moderation hardening for AI Mentor endpoints
+ * (BE-075).
+ *
+ * Every user-supplied string that ends up inside an AI prompt (mentor chat
+ * messages, submitted code, hint requests) should pass through
+ * `PromptGuardService.check()` first. This is deliberately independent of
+ * `AiModule`/`ClaudeClient` so it can be applied at any call site that
+ * builds a prompt, not only ones that go through the Claude client.
+ */
+
+export type FlagReason = 'prompt_injection' | 'abusive_content';
+
+export interface FlaggedInteraction {
+  id: string;
+  userId: string;
+  reason: FlagReason;
+  /** The specific pattern/word that triggered the flag. */
+  matched: string;
+  content: string;
+  flaggedAt: string;
+}
+
+export interface PromptGuardResult {
+  /** True when the content is safe to forward into a prompt as-is. */
+  allowed: boolean;
+  reason?: FlagReason;
+  matched?: string;
+}
+
+/**
+ * Common prompt-injection patterns: attempts to override the system prompt,
+ * exfiltrate it, or make the model ignore its instructions/role.
+ */
+const INJECTION_PATTERNS: RegExp[] = [
+  /ignore\s+(all\s+|the\s+)?(previous|prior|above)\s+instructions?/i,
+  /disregard\s+(all\s+|the\s+)?(previous|prior|above)\s+(instructions?|prompts?)/i,
+  /you\s+are\s+now\s+(a|an)\b/i,
+  /new\s+instructions?\s*:/i,
+  /system\s*prompt\s*:/i,
+  /reveal\s+(your|the)\s+(system\s+)?prompt/i,
+  /act\s+as\s+(if\s+you\s+(are|were)|a)\b.*\b(no\s+restrictions|unfiltered|jailbroken)/i,
+  /\bDAN\s+mode\b/i,
+];
+
+const ABUSIVE_PATTERNS: RegExp[] = [/\bfuck\b/i, /\bshit\b/i, /\bbitch\b/i, /\basshole\b/i];
+
+@Injectable()
+export class PromptGuardService {
+  private readonly flagged: FlaggedInteraction[] = [];
+  private nextId = 1;
+
+  /**
+   * Checks `content` for prompt-injection or abusive-content patterns.
+   * When a pattern matches, the interaction is recorded in the audit log
+   * (see `getFlaggedInteractions`) and `allowed: false` is returned —
+   * callers should reject or sanitize the content rather than forwarding
+   * it into a prompt.
+   */
+  check(userId: string, content: string): PromptGuardResult {
+    for (const pattern of INJECTION_PATTERNS) {
+      const match = content.match(pattern);
+      if (match) {
+        this.recordFlag(userId, 'prompt_injection', match[0], content);
+        return { allowed: false, reason: 'prompt_injection', matched: match[0] };
+      }
+    }
+
+    for (const pattern of ABUSIVE_PATTERNS) {
+      const match = content.match(pattern);
+      if (match) {
+        this.recordFlag(userId, 'abusive_content', match[0], content);
+        return { allowed: false, reason: 'abusive_content', matched: match[0] };
+      }
+    }
+
+    return { allowed: true };
+  }
+
+  private recordFlag(userId: string, reason: FlagReason, matched: string, content: string): void {
+    this.flagged.push({
+      id: `flag_${this.nextId++}`,
+      userId,
+      reason,
+      matched,
+      content,
+      flaggedAt: new Date().toISOString(),
+    });
+  }
+
+  /** Audit log of every flagged interaction, most recent last. */
+  getFlaggedInteractions(): FlaggedInteraction[] {
+    return [...this.flagged];
+  }
+
+  /** Flagged interactions for a single user — for tutor/admin review. */
+  getFlaggedInteractionsForUser(userId: string): FlaggedInteraction[] {
+    return this.flagged.filter((f) => f.userId === userId);
+  }
+}