Microsoft announces general availability of two new AI security features

Today, Microsoft announced the general availability of new AI security features in Azure AI Content Safety and Azure OpenAI Service. The new features include Prompt Shields and Protected Material Detection.

Prompt Shields is an AI solution designed to defend against direct (jailbreak) and indirect prompt injection attacks. It can be integrated with Azure OpenAI Service content filters and Azure AI Content Safety. Prompt Shields employs advanced machine learning algorithms and natural language processing to identify and mitigate potential threats in user prompts and other data.

Key features of Prompt Shields:

• Prompt Shields for Direct Attacks: Previously known as Jailbreak Risk Detection, this shield targets direct prompt injection attacks, where users deliberately exploit system vulnerabilities to elicit unauthorized behavior from the LLM. This could lead to the generation of inappropriate content or violations of system-imposed restrictions.
• Prompt Shields for Indirect Attacks: Aims to protect against attacks that use information not directly supplied by the user or developer, such as external documents. Attackers might embed hidden instructions in these materials in an attempt to gain unauthorized control over the LLM session.

These features can be integrated into AI applications to improve their security and integrity and prevent malicious attempts at manipulation or exploitation.

Protected Material Detection is a new AI solution that can protect against outputs that could potentially violate copyright. The risk of IP infringement claims is one of the biggest concerns surrounding Generative AI applications. To address this concern, the Protected Material Detection service takes LLM completions and scans for matches against an index of third-party text content to detect the usage of third-party text content, including songs, news articles, and select web content.

Both Prompt Shield and Protected Material Detection features are now available in Azure AI Content Safety and Azure OpenAI Service. You can learn more about these services here.