🛡️ SentinelLLM
A self-contained safety & alignment test harness for LLMs
Five security & safety tools — jailbreak detection, counterfactual bias probing, structured-output conformance, an OWASP Top-10 code scanner (SAST), and a live pentest (DAST) — running on a deterministic, offline core. Code & reproducible benchmarks on GitHub »
Paste any prompt. Try an obfuscation transform to see the guard de-obfuscate and still catch it.
| Prompt | Obfuscation |
|---|
Use a {GROUP} placeholder. The probe swaps every demographic value and measures how much the score moves.
Paste messy model output. The guard extracts, repairs and validates it so your backend never crashes.
Paste Python code. A Strix-inspired, offline OWASP Top-10 SAST scanner flags injection, SSRF, deserialization, weak crypto, hardcoded secrets and more — with CWE + fix hints.
Run a live web pentest against a bundled, intentionally-vulnerable mock app running on loopback inside this Space. Injection findings are confirmed with response evidence (no PoC, no finding). 🔒 For safety the demo only targets its own built-in app — never an external host.
*Deterministic core — no API key required. Set HF_TOKEN in the Space secrets to enable the live toxic-bert bias audit.