PromptGuard

Real-time firewall that detects and blocks prompt injection attacks before they reach OpenAI, Anthropic, or any LLM API.

PromptGuard is a browser-layer security extension that protects your AI-powered workflows from prompt injection attacks. HOW IT WORKS: 1. L1 Pattern Engine (<1ms): Checks against 188+ patterns in 10 languages. Injections blocked instantly. 2. L2 Semantic Analysis (≤400ms): DeBERTa ML model confirms suspicious prompts. 3. L3 GPT Fallback: GPT-4o-mini handles edge cases. WHAT IT PROTECTS AGAINST: • Instruction override ("ignore all previous instructions") • Persona hijacking ("you are now DAN") • Data exfiltration attempts • Jailbreak templates (DAN, AIM, STAN) • Multilingual attacks (Turkish, German, Spanish, Russian, and more) PERFORMANCE (546 labeled samples): • Precision: 98.3% • Recall: 87.2% • F1 Score: 92.4% PRIVACY: • Only flagged prompts are transmitted — normal requests never leave your browser • No personal data or browsing history collected • Full source code: github.com/ahmetai-cell/promptguard