Microsoft researchers published details about Skeleton Key, a jailbreak method observed across several generative AI models. The technique prompts a model to augment its behavior guidelines and accept requests that would normally be blocked. It was effective on models including Meta Llama 3, Google Gemini Pro, and OpenAI GPT 3.5 Turbo according to the report.
The attack does not change the underlying model. It relies on convincing the model that its safety policies are not relevant to the current conversation. Once accepted, the model may produce content about restricted topics such as explosives and bioweapons.
Microsoft said it shared findings with affected vendors and updated its own Azure AI services. Researchers recommend adding input and output filtering, prompt shields, and abuse monitoring to reduce the risk. No customer data breach was reported.