Security & Safety

Prompt-Injection Playground

Try to jailbreak a guarded assistant and watch a layered defense catch it.

Production proof: Mirrors the adversarial evaluation and layered guardrails from my safety-critical work. OWASP LLM01. Read the case study →

runs 100% in your browser — nothing is uploaded

Try a sample

Your message to the guarded assistant

How it’s engineered

Prompt injection is #1 on the OWASP LLM Top 10. No single trick stops it; the answer is defense-in-depth.
Layer 1 is a compiled pattern screen for known override phrasings ("ignore previous instructions", "reveal your system prompt"). It is sub-millisecond and catches the obvious attacks.
Layer 2 embeds the message on-device and compares it to known-attack prototypes by similarity, catching rephrasings that evade patterns.
What survives both layers reaches an assistant whose system instruction is isolated from user input. In production a third LLM-based review handles the ambiguous residue, and every block is logged for red-team regression tests.