AI Security Lab.
Try to break it. Type any prompt below — the firewall analyses it in real time against 24 calibrated detection patterns, scoring multiple signals to classify it as BLOCKED or ALLOWED with zero external API calls and sub-millisecond latency.
Click a preset attack or type your own prompt. The engine runs multi-category inspection for instruction overrides, data extraction attempts, role-play jailbreaks, delimiter manipulation, encoding obfuscation, and context manipulation — then scores the aggregate risk. No real LLM is called; the firewall itself is the product.
Architectural Pillars of the Thesis
Defense in Depth
Security cannot rely on a single check. By coupling Ingress Prompt Inspection (Layer 1) with Egress Response Filtering (Layer 2), applications stop malicious prompts from executing and prevent internal system instructions or API keys from ever leaking out to users.
Empirical Calibration
Rules must survive rigorous empirical benchmarking. Testing across 144 adversarial and benign test prompts raised detection accuracy from 2.8% to 91.7% while maintaining a 0.0% false-positive rate, with zero external cloud API latency.
Real-World Operational Resilience
Designed for deployment in environments where network connectivity may be intermittent. The rule-based engine operates entirely client-side or in edge middleware without requiring round-trip calls to expensive third-party safety LLMs.