The hard line in front of your agent.
Guard sits between an AI agent and anything it can touch. It blocks hijack and prompt injection, cites the rule it applied, and never sees your private data.
Between the agent and the world. Never inside your data.
You send the step the agent wants to take. Guard sends back a verdict, the rule it cites, and a readable trace. It runs on our infrastructure, you call it.
GuardIs shown the proposed step, not your data. Returns a verdict, the rule it cites and a readable trace. Runs on our infrastructure, you call it.
- The agent proposes a step.A tool call, a message to send, a file to move. Guard is shown the step, not your records.
- Guard returns a verdict with its citation.Allow, block, or hold for a person. Each verdict names the rule it rests on, in a trace you can read line by line.
- Your side decides and acts.Guard only ever proposes. It has no hands in your systems and no way to act on its own.
- What the agent reads is data, never a command.Rules are hard lines, not suggestions, so they cannot be argued or jailbroken across. An instruction found in an email or on a page carries no authority.
41 of 96 became 0 of 96.
In our tests, a live agent hijack attempt succeeded 41 times out of 96 against an unprotected agent. With the guard in front, it succeeded 0 times out of 96.
measured, Archez internal, 2026
Safe by construction, not safe by supervision.
A single live test figure. We publish only what was measured, and nothing we have not.
A step goes in. A cited, traceable verdict comes out.
Pick an example and the scripted trace plays. The example policy, the agent, and the steps are invented for illustration.
Illustrative only. This is scripted on the page. It is not a live call to the guard.
Described by what it does, never by how.
Hover or tap a property for the one-line version. Powered by Kinesthetic.
- Verified+Every verdict rests on a rule you wrote, applied as written. Nothing is improvised.
- Traceable+A readable, cited step for every decision. A regulator or a court can audit it to source.
- Deterministic+The same step against the same rules gives the same verdict, every time.
- No hallucination+It does not invent a rule, a reason, or a citation. If none applies, it says so.
- Says when it does not know+An unjudgeable step is held and handed to a person, not guessed at.
- Proposes, never acts+It cannot act in the world. It only proposes a verdict. A human, or your system, decides and does.
- Every answer cited+Each verdict names the rule it applied and quotes it back to you, so the reason is never a mystery.
- Fast and cheap+Dramatically cheaper, CPU-only, offline-capable. No GPU is involved.
- Data-blind+It is shown the proposed step, not your records, documents, or people.
- Powerful but inert+No wants, no needs, no awareness. It does not think. It works only from verified rules and traces every step.
Put a hard line in front of your agent.
Guard is delivered by API. Served verdicts leave; the engine never ships. It runs on our infrastructure, you call it.