[2026.3.1-beta]
Fifth Public Beta Release
The AI Security Gateway is a unified security platform providing real-time monitoring, policy enforcement, and threat detection for Large Language Model (LLM) APIs and Model Context Protocol (MCP) servers. This release introduces Guardrail Providers for real-time third-party content screening, a redesigned navigation layout, and a new visual Pipeline view for OAuth access management.
🛡️ Guardrail Providers
A new integration layer that connects the AI Security Gateway to third-party guardrail services for real-time content screening of LLM requests and responses. Multiple providers can run concurrently using a fan-out/fan-in pattern, where total latency equals the slowest provider rather than the sum.
Supported providers:
- Groq Safeguard : High-speed safety classification with configurable safety policy prompts
- EnkryptAI : Comprehensive guardrail API with policy-based detection across NSFW, toxicity, PII, injection attacks, and more
- DynamoAI DynamoGuard : Multi-policy moderation with per-policy scoring for prompt injection, toxicity, PII, hate speech, and violence
- GuardrailsAI : Self-hosted, open-source guardrail with 67+ validators from Guardrails Hub covering jailbreak detection, PII, toxicity, and content policy
- Fiddler AI Guardrails : Sub-second safety classification across 11 dimensions with optional 24-type PII detection
Key features:
- Per-Proxy & Per-Team Assignment: Apply providers globally to a proxy or scope them to specific teams for layered screening
- Configurable Behaviour: Set direction (request/response/both), action (block/monitor), failure mode (fail-open/fail-closed), timeout (500ms–60s), and priority
- Health Checks & Test Playground: Verify provider connectivity and test content screening directly from the dashboard
- Audit Logging: Every provider check is logged with verdict, categories, confidence score, latency, and tokens used
- Dashboard & Metrics: Per-provider and per-proxy statistics with top violation category breakdowns
🧪 Guardrail Evaluation Enhancements
Guardrail Evaluations can now target configured Guardrail Providers directly, in addition to HTTP endpoints. This makes it easy to benchmark provider detection rates against the built-in 71 test cases across 12 attack categories without needing to set up a separate endpoint.
🔔 Granular Alert Notifications
Email and Slack integrations now support fine-grained control over which alerts trigger notifications:
- Per-Category Filtering: Enable or disable notifications for specific alert categories : policy violations, tool access attempts, security events, system events, and compliance alerts
- Minimum Severity Threshold: Set a minimum severity level (Critical, High, Medium, Low, Info) per integration channel
- Each integration channel can be configured independently
🧭 Redesigned Navigation
The main sidebar navigation has been reorganised for improved clarity as the platform grows:
- Security Tools: Hosts System Prompts, Canary Detection, and the Skill Security Hub
- Guardrails: New menu group for Guardrail Providers and Guardrail Evaluations
- Access & Identity: Teams & API Keys, OAuth Access, OAuth Proxy, User Activity, and Playground
- Infrastructure: Proxy Management, A2A Agents, and Audit Logs
🔀 OAuth Access Pipeline View
A new visual Pipeline view for OAuth Access Management that shows the full user journey from provider to permissions:
Provider → Rules → Teams → Users → Tool Overrides → Summary
Each stage is an interactive section where administrators can drill into provider configurations, group assignment rules, team memberships, individual users, and per-user tool permission overrides. A summary bar displays quick statistics across the pipeline. The view mode (Pipeline or Table) is persisted across sessions.
🐛 Bug Fixes & Improvements
- Improvements to Canary Token Detection reliability and accuracy
- System Prompt Injection stability fixes
- External Storage sync and configuration improvements
- General UI polish and consistency updates