Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Learn what's new in the service. These items might be release notes, videos, blog posts, and other types of information. Bookmark this page to stay up to date with new features, enhancements, fixes, and documentation updates. October 2026 Content Safety Model Update
October 2026
We are introducing an update to text content harm classification (hate, sexual, self-harm, and violence) and Prompt Shields detection of jailbreak (user prompt injection attacks, UPIA) and indirect prompt injection attacks (XPIA) in Azure AI Content Safety and Microsoft Foundry Guardrails. The update is intended to address reported classification issues, expand multilingual training coverage, and improve the balance between detecting content harms and avoiding unnecessary filtering.
This release doesn't update the behavior of personally identifiable information (PII) detection, task adherence, network egress controls, Spotlighting, block lists, protected material detection for code or text, or image safety capabilities.
What is changing
Expanded training coverage for detection of jailbreak and indirect prompt injection attacks in Chinese, French, German, Italian, and Portuguese, intended to help distinguish attacks from ordinary content in these languages.
Relabeled training data for hate, sexual, self-harm, and violence by using an updated risk taxonomy, intended to support clearer distinctions between classes.
Targeted training updates informed by reported classification issues, intended to improve how the model handles similar scenarios.
Adjustments to classification sensitivity, intended to better balance detection of content harms with avoidance of false detections.
Expanded evaluation coverage to help identify potential classification issues and assess rollout readiness.
What customers might observe
Some text content might receive different classifications, severity levels, or filtering outcomes for text content harms and prompt injections.
The effects of this update might vary by input, language, content category, and configured filtering settings. These changes don't guarantee that every reported issue is resolved or that every scenario improves.
Rollout and availability
This update is now available in Spain Central, the initial canary production region for this rollout. Test representative application scenarios in this region and share your feedback.
Expansion will proceed in stages, subject to validation and review of rollout feedback. Timing might change as rollout progresses.
During rollout, behavior might temporarily differ across regions. We will update this notice as rollout progresses and announce when the overall rollout is complete.
Customer action and feedback
Test representative content by using an Azure AI Content Safety resource or a Foundry resource in Spain Central. Review your filtering configuration and monitor classification and blocking behavior as the update reaches your deployments.
We encourage all customers to share what is working well and what could improve through this feedback form, whether or not you have an Azure technical support plan. A brief observation is welcome; technical details and examples are optional. The model team is actively reviewing submissions during rollout to assess model behavior and investigate issues.
If you choose to share an example, include the result you received and what you expected. If you prefer to share an example as a file, you can email contentsafetysupport@microsoft.com with the subject "Content safety feedback".
Remove secrets, credentials, personal information, and confidential content from feedback text, examples, and files. If you can't safely share an example, describe the scenario instead. You can optionally provide a contact email in the form if you're open to follow-up.
This form collects product feedback, not technical support requests. Submitting it doesn't open a support case or guarantee an individual response or resolution. If you need technical support, including help with a production issue, use your usual support channel.
Upcoming deprecations
The following Azure AI Content Safety preview capabilities retire on April 30, 2027. Review the feature-specific migration guidance and update affected workloads before that date.
- Analyze Multimodal.
- Custom categories (rapid), including text and image.
- Connected Content Safety containers.
- Disconnected Content Safety containers.
- Embedded Content Safety SDK.
- Groundedness detection in Azure AI Content Safety and Azure OpenAI.
November 2025
Task Adherence public preview
The Task Adherence feature identifies discrepancies between an LLM's behavior and its given task, such as misaligned tool invocations, improper tool input or output relative to user intent, and inconsistencies between responses and customer input. This feature lets system developers proactively mitigate misaligned actions by blocking them or escalating the issue for human intervention.
Learn more on the Task Adherence concepts page, and get started with the quickstart.
October 2024
Upcoming deprecations
To align with Content Safety versioning and lifecycle management policies, the following versions are scheduled for deprecation:
- Effective March 1st, 2025: All API versions except
2024-09-01,2024-09-15-preview, and2024-09-30-previewwill be deprecated and no longer supported. We encourage users to transition to the latest available versions to continue receiving full support and updates. If you have any questions about this process or need assistance with the transition, please reach out to our support team.
September 2024
Multimodal analysis (preview)
The Multimodal API analyzes materials containing both image content and text content to help make applications and services safer from harmful user-generated or AI-generated content. Analyzing an image and its associated text content together can preserve context and provide a more comprehensive understanding of the content. Follow the quickstart to get started.
Protected material detection for code (preview)
The Protected material code API flags protected code content (from known GitHub repositories, including software libraries, source code, algorithms, and other proprietary programming content) that might be output by large language models. Follow the quickstart to get started.
Caution
The Content Safety service's code scanner/indexer is only current through April 6, 2023. Code that was added to GitHub after this date won't be detected. Use your own discretion when using Protected Material for Code to detect recent bodies of code.
Groundedness correction (preview)
The groundedness detection API includes a correction feature that automatically corrects any detected ungroundedness in the text based on the provided grounding sources. When the correction feature is enabled, the response includes a corrected Text field that presents the corrected text aligned with the grounding sources. Follow the quickstart to get started.
August 2024
New features are GA
The Prompt Shields API and Protected Material for text API are now generally available (GA). Follow a quickstart to try them out.
July 2024
Custom categories (standard) API public preview
The custom categories (standard) API lets you create and train your own custom content categories and scan text for matches. See Custom categories to learn more.
May 2024
Custom categories (rapid) API public preview
The custom categories (rapid) API lets you quickly define emerging harmful content patterns and scan text and images for matches. See Custom categories to learn more.
March 2024
Prompt Shields public preview
Previously known as Jailbreak risk detection, this updated feature detects prompt attacks, in which users deliberately exploit system vulnerabilities to elicit unauthorized behavior from large language model. Prompt Shields analyzes both direct user prompt attacks and indirect attacks which are embedded in input documents or images. See Prompt Shields to learn more.
Groundedness detection public preview
The Groundedness detection API detects whether the text responses of large language models (LLMs) are grounded in the source materials provided by the users. Ungroundedness describes instances where the LLMs produce information that is non-factual or inaccurate according to what was present in the source materials. See Groundedness detection to learn more.
January 2024
Content Safety SDK GA
The Azure AI Content Safety service is now generally available through the following client library SDKs:
- C#: Package | API reference | Samples | Quickstarts: Text, Image
- Python: Package | API reference | Samples | Quickstarts: Text, Image
- Java: Package | API reference | Samples | Quickstarts: Text, Image
- JavaScript: Package | API reference | Samples | Quickstarts: Text, Image
Important
The public preview versions of the Azure AI Content Safety SDKs will be deprecated by March 31, 2024. Please update your applications to use the GA versions.
November 2023
Jailbreak risk and protected material detection (preview)
The new Jailbreak risk detection and protected material detection APIs let you mitigate some of the risks when using generative AI.
- Jailbreak risk detection scans text for the risk of a jailbreak attack on a Large Language Model. Quickstart
- Protected material text detection scans AI-generated text for known text content (for example, song lyrics, articles, recipes, selected web content). Quickstart
Jailbreak risk and protected material detection are only available in select regions. See Region availability.
October 2023
Azure AI Content Safety is generally available (GA)
The Azure AI Content Safety service is now generally available as a cloud service.
- The service is available in many more Azure regions. See the Overview for a list.
- The return formats of the Analyze APIs have changed. See the Quickstarts for the latest examples.
- The names and return formats of several other APIs have changed. See the Migration guide for a full list of breaking changes. Other guides and quickstarts now reflect the GA version.
Azure AI Content Safety Java and JavaScript SDKs
The Azure AI Content Safety service is now available through Java and JavaScript SDKs. The SDKs are available on Maven and npm respectively. Follow a quickstart to get started.
July 2023
Azure AI Content Safety C# SDK
The Azure AI Content Safety service is now available through a C# SDK. The SDK is available on NuGet. Follow a quickstart to get started.
May 2023
Azure AI Content Safety public preview
Azure AI Content Safety detects material that is potentially offensive, risky, or otherwise undesirable. This service offers state-of-the-art text and image models that detect problematic content. Azure AI Content Safety helps make applications and services safer from harmful user-generated and AI-generated content. Follow a quickstart to get started.