Crazy. All you have to do to get around the censorship of ChatGPT and other AI models is change your prompt very slightly, as little as a single letter. It usually takes multiple attempts, but is effective 89% of the time according the the video.
Sort: Trending
Here's a summary:
Anthropic's groundbreaking Shotgun Jailbreaking cracks all Frontier AI models
Anthropic introduces "Shotgun Jailbreaking," a simple yet highly effective method to bypass restrictions across Frontier AI models, including text, vision, and audio systems. This technique involves generating numerous prompt variations, such as leetspeak, capitalization changes, or audio/visual tweaks, until the model produces the desired output. With success rates as high as 89% for GPT-4 and 78% for Claude 3.5 Sonet, this method is scalable and works well alongside other jailbreak techniques. Anthropics' paper highlights the inevitability of such vulnerabilities in AI models, aiming to raise awareness and improve security. The technique and its code are open-sourced for testing.