Microsoft is including security and safety instruments to Azure AI Studio, the corporate’s cloud-based toolkit for constructing generative AI purposes. The brand new instruments embrace safety towards immediate injection assaults, detection of hallucinations in mannequin output, system messages to steer fashions towards secure output, mannequin security evaluations, and threat and security monitoring.
Microsoft introduced the brand new options on March 28. Security evaluations at the moment are out there in preview in Azure AI Studio. The opposite options are coming quickly, Microsoft stated. Azure AI Studio, additionally in preview, might be accessed from ai.azure.com.
Immediate shields will detect and block injection assaults and embrace a brand new mannequin to establish oblique immediate assaults earlier than they influence the mannequin. This characteristic is at present out there in preview in Azure AI Content material Security. Groundness detection is designed to establish text-based hallucinations, together with minor inaccuracies, in mannequin outputs. This characteristic detects “ungrounded materials” in textual content to assist the standard of LLM outputs, Microsoft stated.
Security system messages, also referred to as metaprompts, steer a mannequin’s conduct towards secure and accountable outputs. Security evaluations assess an utility’s capability to jailbreak assaults and to producing content material dangers. Along with mannequin high quality metrics, they supply metrics associated to content material and safety dangers.
Lastly, threat and security monitoring helps customers perceive what mannequin inputs, outputs, and customers are triggering content material filters to tell mitigation. This characteristic is at present out there in preview in Azure OpenAI Service.
Copyright © 2024 IDG Communications, Inc.
