AI Runtime Guardrails
Check every AI action before it runs.
Relyance Guard sits between your AI and the data, tools, and APIs it can reach. Every call is allowed, masked, or blocked in milliseconds, and logged for review.
What you get
Every AI app, checked on every call.
Any AI feature you ship gets the same treatment: the prompt, the retrieved context, and the response are all checked before they move in your ecosystem.
Autonomous agents stay inside the task & goal
An agent clearing a refund backlog can refund $40 and not $12,400, and instructions hidden inside a ticket never become its new orders. One control point covers every model and framework.
Coding agents and people, same rules.
A coding agent running with approvals off never reads production secrets, and a customer list pasted into a public chat tool stops at the proxy. Every decision is logged with the evidence behind it.
CAPABILITIES
Enforce approved AI actions
Every command and API call is checked against the job it was given, at the moment it runs.
Checked at the call
Each action is resolved against the live data graph and the declared purpose of the task, not just a role.
Blocked in CI
Over-scoped agents fail the pipeline before they ship, with the exact permission that tripped it.
No static allow-lists
The decision is computed when it matters, so a change in data or intent changes the answer.
Block data misuse
Text written by a customer or untrusted input from a source is never treated as an instruction.
Purpose, not just permission
Valid access is flagged the moment it is used outside the purpose it was granted for.
Injection refused inline
Prompt injection, direct or indirect, is stripped from the context before the model acts on it.
Compound risk, scored
Risk is scored across the graph of data, identity, and destination, not permission by permission.
Prevent sensitive data leaks
Sensitive data is stopped before it reaches a prompt, a context window, agent memory, or a log line — even when the model describes or paraphrases it instead of naming it directly.
Stopped before the model
Fields are classified in flight and held back before egress to any provider or third party.
Redact, mask, escalate, or narrow
A hard block is not always the right call. Guard can mask a field or narrow the scope instead or it can also raise a call for human review.
Verifiable evidence
Every decision is logged with the data origin, acting identity, and owner, ready for audit.
HOW IT WORKS
One request. Four facts. One answer.
Guard resolves each call against the live data graph: what the data is, where it came from, which identity is acting, and what the task is meant to do. Then it answers in single-digit milliseconds.
1
Classification and origin of every field in the request
2
Identity and permissions the agent is actually running on
3
Declared purpose of the task and agent blueprint
4
Destination: model provider, tool, memory, log, or third party
Deployment options
Where it runs
As a gateway or proxy in front of model providers, or as a plugin to your existing AI gateway.
What it sees
Payload inspection happens in your environment. Relyance stores decisions and evidence, not your prompts or data.
When it fails
Fail-open or fail-closed is set per policy. Start on one endpoint in observe mode; flip to enforce when the decision log looks right.
Bring one agent. Leave with its first blocked request.
In the demo we point Guard at a service you choose, replay real traffic in observe mode, and walk through every decision it would have made.
30-minute demo
Your agent, your policy, your traffic
Decision log you keep afterwards
Private beta access if it fits