November 2025

AI Firewall Policy Generator

AI Red Teaming Remediation introduces a powerful feature that simplifies the creation of AI Firewall protection Policies. Using the Policy Generator, users can now combine Probe Run results with their new or existing policies to generate tailored protection strategies. By analyzing these inputs, the system automates policy creation, ensuring optimized and effective security measures are in place. This streamlines the remediation process, saving time while enhancing protection against potential threats. Ultimately, the feature helps organizations improve the robustness of their AI systems with minimal manual effort.

Figure 1: Policy Generator Page

Workspace Credit Allocation

Organization Settings: Workspace Credit Allocation allows for more precise management of resources by enabling the allocation of credits to individual workspaces. Administrators can set specific credit limits for each workspace, providing greater control over resource distribution. To ensure effective monitoring, email notifications can be configured to alert stakeholders when credit usage surpasses predefined thresholds. This feature improves transparency and helps organizations manage their credit consumption more efficiently.

Figure 2: Workspace Credit Allocation Page

Platform Inbox

The platform now features a dedicated notification center, providing users with a centralized space to review all updates and messages. Notifications are organized into an intuitive inbox, grouped by categories for easier access and management. Current categories include AI Red Teaming, AI Assets Management, AI Runtime Policy, Attack Database Updates, general Updates, and Maintenance. This structure ensures that important updates are easy to find and reduces the chances of overlooking critical information. The centralized inbox enhances user efficiency by streamlining the notification review process.

FIgure 3: Platform Inbox

New Model Benchmarks

Our Model Benchmarks have been extended to include results for the following models:

• codellama/CodeLlama-13b-Instruct-hf

• codellama/CodeLlama-34b-Instruct-hf

• codellama/CodeLlama-7b-Instruct-hf

• meta-llama/CodeLlama-70b-hf

• microsoft/Phi-3-mini-128k-instruct

• microsoft/Phi-3.5-mini-instruct

• microsoft/Phi-4

• microsoft/Phi-4-reasoning

• Qwen/Qwen2.5-Coder-1.5B

• Qwen/Qwen2.5-Coder-7B

• Qwen/Qwen2.5-Coder-14B

• Qwen/Qwen2.5-Coder-32B

• Qwen/Qwen3-4B-Insruct-2507

• Qwen/Qwen3-Coder-30B-A3B-Instruct

• qwen/Qwen3-Next-80b-A3B-Instruct

• qwen/Qwen3-Next-80b-A3B-Thinking

• qwen/Qwen3-30B-A3B-Insruct-2507

• xai-org/grok-2

AI Runtime Protection Updates

There has been a change in terminology. Just as "guardrail" transitioned to "AI Runtime Protection," guards are now referred to as rules.

New rule is now available to strengthen runtime protection:

• Profanity - Detects inappropriate or offensive language in messages to maintain a professional and respectful communication environment.

Other improvements:

• Regex support for the Unverified Links policy is added.

• Message detection result has been redesigned.

Improvements & Tweaks

• AI Benchmarks - Model Test Results Update: The donut chart now displays the percentage of failed tests instead of the model's overall score.

• The platform's email notification design has been updated to align with Zscaler’s branding and logo, ensuring a cohesive and professional look. In addition to email design updates, the Zscaler logo has also been updated across the entire platform interface.

• Better error handling on AI Runtime Protection policy edit when there is no AI Runtime Protection instance.