Changelog

API Updates

• Guard API: The breakdown response now includes a result field for each detector, showing the confidence level (l1_confident, l2_very_likely, l3_likely, l4_less_likely, l5_unlikely, or no_level). This provides the same granular confidence information available in the /guard/results endpoint, allowing you to see not just whether a detector flagged content, but also how confident the detection was.

• Guard API: A new sub category self-harm is added to the content moderator detector.

New Features

• Dashboard: Policy Impact Simulator - an interactive tool that shows how different sensitivity levels and guardrail configurations would have affected your historical traffic. Available on policy view and edit pages, as well as in the policies list page as a column. Compare flagging rates across L1-L4 and see category-level breakdowns to tune your policies with confidence.

New Features

• Guard: For audio requests, the audio_payload flag in the request provides access to debugging information of the sample.
• Gateway: For audio requests, sensitive audio material is no longer included in logs.

Quality

• General: Model improvements based on client feedback.
• Audio (TensorRT-LLM): Added denoising to audio processing.

Bug Fixes

• Guard: Fixed a bug in policy handling for audio requests where L2 samples were mistakenly labeled as L3.
• Gateway: Request metadata is no longer altered for logging purposes. User-supplied metadata is now treated strictly as provided.

New Features

• Gateway: Also log requests to v2/guard/audio for monitoring purposes.

Quality

• Prompt injection: Improved detection of malicious behavioral instructions.
• Prompt injection: Improved detection of system prompt exfiltration.
• Prompt injection: Retrained model with improved coverage on new prompt attack variants.
• Content moderation: Model update to include customer feedback.

Bug Fixes

• Guard: Health probe fixes when gRPC encryption is turned on.
• Gateway: Fix policy resolution for multi-message requests.

Security

• Gateway: Security fixes (CVE-2026-25679, CVE-2026-27142, CVE-2026-27139).

Quality

• Improved GPU model for text and audio.
• Reduction of FPRs in audio guard.
• General model quality improvement based on customer feedback.

Quality

• Updated prompt injection models incorporating customer feedback for improved accuracy.
• Fixed bug related to PII spans. Also improved our performance on detecting credit card numbers.
• Improved latency and throughput for self-hosted deployments.

Quality

• Fixed an issue where allowlists could be incorrect.
• Updated moderation models incorporating customer feedback for improved accuracy.

Improvements

Quality

• Moderation model improvements: Updated moderation models to reduce FPRs, especially in weapons category.
• Text preprocessing robustness: Improved handling of escaped JSON characters and edge cases in text decoding, reducing preprocessing errors and improving classifier reliability.
• Whitelist refinements: Removed common phrases from whitelist to improve detection accuracy.

Improvements

Platform

• Fixed bug breaking onboarding page for some users.

Quality

• Expanded API to cover tool role and tool_calls to better support Agentic workflows.

Improvements

Platform

• Improve Logs page loading speed.
• Fixed bug which disallowed the same detector to be used for input and output.

Quality

• Improved handling base64-encoded input.