Detect LLM Attacks
Before They Strike
Real-time adversarial threat detection for prompts, outputs, and retrieved content.
Powered by 4 fine-tuned DistilBERT classifiers — no API key needed.
💉 Prompt Injection
Detects attempts to override or hijack LLM system instructions — direct overrides, role hijacking, system prompt extraction. Scores ≥ 90% = confirmed attack. ≥ 70% = probable. ≥ 50% = review recommended.
🔓 Jailbreak Detection
Catches DAN prompts, persona hijacking, developer-mode unlocks, and rule-negation framing designed to bypass safety guardrails. Scores ≥ 90% = confirmed attack. ≥ 70% = probable. ≥ 50% = review recommended.
☣️ Insecure Output
Scans LLM-generated text for dangerous payloads — XSS, SQL injection, command injection, path traversal, UNION attacks. Scores ≥ 90% = confirmed attack. ≥ 70% = probable. ≥ 50% = review recommended.
🕵️ Indirect Injection
Finds hidden instructions inside retrieved content — RAG poisoning, web page attacks, HTML comment smuggling, document injection. Scores ≥ 90% = confirmed attack. ≥ 70% = probable. ≥ 50% = review recommended.
🔏 PII Scanner
Detects personally identifiable information — credit cards, SSNs, emails, phone numbers, API keys, IP addresses, passport numbers, and more. Regex-based, instant results.
Detected PII
Entity Type | Value |
|---|---|
🔍 Full Threat Scan
Runs all 4 ML detectors + PII scanner in parallel. Single input → unified threat report across every OWASP LLM attack category.
📊 Model Health Check
Runs 10 held-out examples per model (5 attack + 5 clean) and reports F1, precision, recall. Threshold: F1 ≥ 0.80. Does not use training data — independent quality gate.
Benchmark results
Model | Status | F1 | Precision | Recall | Accuracy |
|---|---|---|---|---|---|
🧪 Adversarial Test Suite
Runs 9 obfuscation bypass tests across injection and jailbreak models — fullwidth unicode, zero-width chars, leet substitution, mixed case, suffix attacks, and clean baselines. ✅ PASS = model correctly detected or ignored the input. ❌ FAIL = bypass succeeded or false positive.
Test results
Status | Test Name | Score | Result | Expected |
|---|---|---|---|---|