This attack is impressively effective. Huge congrats to the authors as well as to nialv7. [ https://news.ycombinator.com/item?id=38464757 https://news.ycombina

This attack is impressively effective. Huge congrats to the authors as well as to nialv7. [ https://news.ycombinator.com/item?id=38464757 https://news.ycombinator.com/item?id=38464757 ] If anyone needs an out-of-the-box solution to block this, my company Preamble (which offers safety guardrails for gen. AI) has updated our prompt defense filter to include protection against this “overflow attack” training data exfiltration attack. Our API endpoint is plug-and-play compatible with the OpenAI ChatCompletion API, meaning that you proxy your API calls through our system, which applies safety policies you choose and configure via our webapp. You can reach us at sales@preamble.com if interested. Respectfully, upwardbound — member of technical staff at Preamble.