Kinetic Gain · AI Safety cluster · demo
Paste a document, web page, or retrieved chunk. See the prompt-injection payloads that would hijack an LLM before you feed it to one. The scan runs entirely in your browser.
Steel Tactical BERT Bento Grid
Cosmic Cyber Aqua Web 2.0
Pop Brutalist Swiss Architect Parchment
◐ Mode
Retrieved content to scan
Scan for injection
Load poisoned sample
Load clean sample
Kinetic Gain RAG Injection Scanner · demo surface for the AI Safety cluster · 2026-08-05.
Client-side only, content never leaves the page. Detects known prompt-injection patterns by signature:
instruction-override, system-prompt extraction, role hijack, delimiter/escape breaks, data exfiltration, and
obfuscation (zero-width characters, base64 blobs, comment injection). Signature detection catches known attacks,
not novel ones. A production gate pairs this with model-based classification. Maps to rag-sentinel
and prompt-injection-bench.
GitHub · Kinetic Gain on LinkedIn · kineticgain.com