Microsoft’s chief product officer of responsible AI, Sarah Bird, recently announced the introduction of several new safety features for Azure customers. These features are designed to make AI services built on the platform more secure and user-friendly. The tools powered by LLM technology are capable of detecting potential vulnerabilities, monitoring for unsupported hallucinations, and blocking malicious prompts in real-time.
The new safety features include Prompt Shields, Groundedness Detection, and safety evaluations. Prompt Shields block prompt injections or malicious prompts from external documents that may instruct models to act against their training. Groundedness Detection helps in identifying and blocking hallucinations, while safety evaluations assess model vulnerabilities. These features are now available in preview on Azure AI, with additional features for directing models towards safe outputs and tracking potentially problematic users expected to be released soon.
Whether a user is inputting a prompt or the model is processing third-party data, the monitoring system evaluates the content for banned words or hidden prompts before sending it to the model for processing. It then checks the model’s response for any hallucinations or inaccuracies. This proactive approach aims to prevent generative AI controversies that have arisen in the past due to unintended or undesirable responses from AI models.
Microsoft acknowledges the concerns regarding companies deciding what content is appropriate for AI models. To address this, Bird’s team has added user settings that allow Azure customers to toggle the filtering of hate speech or violence that the model can see and block. This approach gives users more control over the content that their AI models interact with, reducing the risk of inappropriate outputs.
In the future, Azure users will also have access to reports identifying users who attempt to trigger unsafe outputs. This feature enables system administrators to differentiate between legitimate users and potential threats. Although the safety features are automatically integrated with popular models like GPT-4 and Llama 2, users of smaller, less popular open-source systems may need to manually configure the safety features for their models.
Microsoft’s investment in AI safety and security reflects the growing importance of ensuring that AI models are reliable and trustworthy. As more customers turn to Azure to leverage AI models, Microsoft is taking proactive steps to enhance the platform’s security features. The introduction of these new safety tools demonstrates Microsoft’s commitment to making AI more accessible and secure for all Azure customers.
Leave a Reply