How they work
- Open AI Gateway, then Guardrails. The page says: “These check every key that is not on an AI system. A key on an AI system is checked by that AI system’s guardrails only.” It has the same eight guardrails as an AI system. See Guardrail types.
- It works the way AI system guardrails do: a draft, Save draft, Apply as a numbered version, Test guardrails and Roll back under Versions. There is no Risks and controls tab, because the organization has no assessment.
- Requests made with an organization key or a personal key are checked against the organization’s live version.
- Requests made with an AI system’s key are checked against that AI system’s live version instead. If the AI system has none, its requests are not checked: the organization’s guardrails do not apply to it. The AI system’s AI Gateway tab says so: “No guardrails are live for this AI system, so nothing checks its keys. The organization’s guardrails do not apply to them.”
- A key that nothing checks shows Unguarded next to Active in Virtual keys and on the AI system’s AI Gateway tab.
- When an organization key is attached to an AI system, its next request uses the AI system’s guardrails.
- Applying a version sends it to the AI Gateway. If the AI Gateway cannot be reached at that moment, the version is still live in Altrum, the screen says “Version 3 is live. Not yet sent to the AI Gateway. Altrum keeps trying.” (with the version number), and Altrum keeps trying until it has.