Skip to content
Halo by OmniTrust

Secure · Runtime Guardrails

Available now· Selected detectors in supported gateway paths

Check prompts and responses at the gateway

Halo runs selected classifiers on traffic that passes through its gateways, looking for PII, secrets, prompt injection, moderation issues, and language. Policy decides whether to allow, mask, redact, hold, or block the content.

Book a walkthrough Start with Halo
check types, from PII to language
5
to test a rule before enforcing it
Audit mode

Why it matters

Prompts and responses carry sensitive content

People paste customer data into prompts, tools return records a model should not repeat, and documents can hide instructions that try to redirect an agent. Access rules alone do not see what is inside a request.

  1. 01

    Prompts can contain personal data or secrets that should not reach a provider.

  2. 02

    Retrieved content can carry hidden instructions that try to redirect an agent.

  3. 03

    Responses can echo sensitive values back to a user or to another system.

How it works

How a guardrail rule works

Each step keeps a record, so you can tune a rule with evidence.

  1. 1

    Write the rule

    A policy rule names the detector, whether it applies to the request or the response, and the outcome, for example redact on a PII result.

  2. 2

    Test in audit mode

    Run the rule in audit mode first. Halo records matches without changing the outcome, so you can check for false positives.

  3. 3

    Enforce at the gateway

    In enforce mode, the gateway applies the result by masking, redacting, holding for approval, or denying the request or the response.

  4. 4

    Review and tune

    Detections and outcomes appear in the evidence trail with the rule that fired, so reviewers can adjust thresholds and conditions.

Capabilities

Checks that run on gateway traffic

Detectors run in supported gateway paths on requests and responses, and policy decides what happens next.

  • PII detection

    In supported gateway paths, selected detectors look for personal data in requests and responses, so policy can mask, redact, or block.

  • Secret scanning

    Inline checks look for credentials and other secrets in prompts, arguments, and responses before they move on to a model or a tool.

  • Prompt injection checks

    A prompt injection detector evaluates incoming content, so policy can flag or block instructions that try to redirect an agent.

  • Moderation and language

    Moderation and language checks give policy more signals, for example content that needs review or text in a language an agent should not handle.

  • Masking and redaction

    Where supported, Halo can mask or redact sensitive values in a payload before it continues to the model or the tool.

  • Policy-driven outcomes

    Policy chooses what happens after a detection, such as allow, mask, hold for approval, or deny, and the outcome is recorded as evidence.

Where checks run

Clear on where checks apply

Detectors run on traffic that passes through supported gateway paths. Halo names the detector types it offers and applies them where the gateway can inspect the payload, so your team knows which traffic a check covers and which it does not.

  • Selected detectors for PII and secrets
  • Prompt injection, moderation, and language checks
  • Masking and redaction in supported paths
  • Audit mode to test before enforcing

FAQ

Questions, answered

Halo offers selected runtime classifiers for personal data (PII), secrets, prompt injection, moderation, and language. They run in supported gateway paths, and availability can differ by path, so confirm the detectors you need during a walkthrough.

In supported gateway paths, policy can mask or redact sensitive values in a payload before it continues. Redaction depends on the detector and the path, so it does not apply to traffic that never passes through a supported gateway path.

Classifier results are probabilistic, so a rule can match content that is harmless. Start rules in audit mode, review the matches in the evidence trail, adjust thresholds or conditions, and move to enforce when the results look right.

Browser DLP checks are separate. Halo Browser applies regex-based checks on supported AI sites, and content inspection needs expanded permissions that your organization approves. Runtime guardrails apply to traffic that passes through Halo gateways.

No detector offers a guarantee. A prompt injection detector adds a check on incoming content, and it works alongside tool permissions, approvals, default-deny egress, and credential brokering that limit what a redirected agent can do.

Related capabilities

Model GatewayManaged and BYOK model accessLearn more about Model GatewayMCP and Tool GatewayGovern MCP servers and tool callsLearn more about MCP and Tool GatewayPolicy EnginePolicy as code with observe and enforceLearn more about Policy Engine

Put a halo around the AI your business runs.

See discovery, runtime policy, approvals, and evidence working together on your own use cases.

Book a walkthrough Start with Halo