<ul><li>Hackers are using artificial intelligence to attack AI models using a technique called Fun-Tuning.</li><li>Fun-Tuning makes prompt injection attacks more effective and achieved up to 82% success rates on Google's Gemini models.</li><li>These attacks exploit subtle clues in the fine-tuning process to increase the chances of successful prompt injections.</li><li>Defending against these attacks is challenging and removing key data from training would make the tool less useful for developers.</li></ul>

Hackers are now using AI to break AI – and it’s working

Discover more