Bifrost Guardrails
AI model · Safety / moderation · Text → Tables / structured data by Maxim AI
Real-time AI security and compliance for enterprise LLM traffic.
Bifrost Guardrails is a self-hosted AI security layer that detects PII, blocks prompt injections, and filters unsafe content in real-time. It enables enterprises to enforce custom safety policies within their own VPC using native checks or third-party integrations.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Languages
- English
Category: AI Infrastructure
- Modalities
- Text, Image → Text, Image
- Latency
- In-process parallel execution with no added latency
- License
- Apache License 2.0 (open source core)
- Access
- Self-hosted API gateway, npx, Docker, Kubernetes
- Deployment Format
- Self-hosted (VPC, On-prem, Air-gapped)
- Core Architecture
- Two-tier (Rules + Profiles)
- Real-time PII, PHI, and secrets redaction
- Blocks direct and indirect prompt injections
- Chains native and third-party guardrail providers
- Self-hosted in-VPC deployment for data privacy
- Custom validation logic via Common Expression Language (CEL)
Overview
Bifrost Guardrails is a real-time AI security and moderation layer designed for enterprises to protect their LLM interactions. Integrated directly into the Bifrost Gateway, it allows organizations to detect and redact PII, block prompt injections, and filter unsafe content before it reaches a model or a user. The system operates within the customer's own infrastructure—whether in-VPC, on-prem, or air-gapped—ensuring that sensitive data is inspected without leaving the private network. It features a flexible two-tier architecture of Rules and Profiles, enabling teams to use native checks or chain third-party providers like AWS Bedrock and Azure Content Safety. By using Common Expression Language (CEL), administrators can write custom validation logic to enforce granular policies across different models and teams.
Detection & Redaction Capabilities
- PII & PHI Redaction: The system redacts emails, SSNs, financial data, and medical records in real-time before they leave the gateway.
- Prompt Injection Blocking: Bifrost blocks direct and indirect prompt injection attacks hidden inside tool results or retrieved content.
- Unsafe Content Filtering: Organizations can block toxic, unsafe, or off-policy content across both text and images.
- Secrets Detection: The platform detects and masks leaked API keys, tokens, and credentials using Gitleaks-backed scanning.
Policy & Rule Management
- Custom CEL Policies: Users can write custom validation logic using Common Expression Language (CEL) to target specific requests by model, role, or content length.
- Provider Chaining: The gateway supports chaining multiple guardrail providers, such as AWS Bedrock and Azure Content Safety, in a single rule.
- Severity Thresholds: Administrators can tune severity thresholds to choose whether to block, redact, or modify content.
- Input/Output Targeting: Rules can be applied selectively to inputs, outputs, or both for layered enforcement.
Governance & Infrastructure
- Guardrail Auditing: Every decision is logged with details on the rule, provider, violation type, severity, and action taken.
- Log Exporting: Guardrail logs can be exported to external destinations including S3, GCS, or BigQuery for long-term storage.
- Native Performance: Native guardrail checks run in-process and in parallel with no added latency to the request path.
- In-VPC Validation: The self-hosted deployment ensures all validation occurs within the customer's perimeter without sending data to third parties.
Use cases
- Security teams redact SSNs and financial data from prompts before they reach external LLM providers.
- Platform engineers block prompt injection attacks hidden in retrieved content to protect agentic workflows.
- Compliance officers maintain immutable audit trails of all AI safety violations for regulatory reviews.
- Developers use native secrets detection to prevent leaked API keys and credentials from being sent to models.
Specifications
- Modalities
- Text, Image → Text, Image
- Latency
- In-process parallel execution with no added latency
- License
- Apache License 2.0 (open source core)
- Access
- Self-hosted API gateway, npx, Docker, Kubernetes
- Deployment Format
- Self-hosted (VPC, On-prem, Air-gapped)
- Core Architecture
- Two-tier (Rules + Profiles)
- Policy Language
- Common Expression Language (CEL)
- Native Detection
- Gitleaks-backed secrets detection, PII templates, Regex
- External Integrations
- AWS, Azure, Google, CrowdStrike, GraySwan, Patronus AI
- Latency Overhead
- 0ms (In-process parallel execution)
- Audit Standards
- Immutable, timestamped records with hash-chaining
- Status Codes
- HTTP 200 (Pass), 446 (Block), 246 (Warning/Redact)
How it works
- Configure Profiles: Configure native or external guardrail providers once, setting credentials and detection thresholds.
- Define Rules: Write CEL logic to target specific requests by model, role, or content length and link them to your profiles.
- Enforce on Traffic: Apply rules to incoming prompts, outgoing model responses, or both for comprehensive protection.
- Audit Decisions: Review immutable, timestamped records of every violation and action in the live request log.
FAQ
What is Bifrost Guardrails?
Bifrost Guardrails is a real-time security layer that sits within the Bifrost Gateway to detect PII, block prompt injections, and enforce safety policies on every LLM request and response.
Which guardrail providers are supported?
Bifrost offers native checks for PII and secrets, and also integrates with external providers like AWS Bedrock Guardrails, Azure Content Safety, and Google Model Armor.
How are guardrail policies configured?
Bifrost uses a two-tier architecture where 'Profiles' configure the provider settings and 'Rules' use Common Expression Language (CEL) to decide when and how to apply those checks.
Added by AntSpace from public sources (getmaxim.ai).