Platform

Automatic responses to AI problems, as far as you allow

Guardian is Aicura's automatic response to problems with AI systems and agents in production. When something goes wrong, it tells someone, waits for an approver or acts straight away, depending on what you have allowed.

Included in Pro and Enterprise. Compare plans

When something goes wrong and nobody is watching

Something changes in production when nobody is looking. An agent starts calling a tool it never declared, or errors climb on a customer-facing system overnight. Guardian notices, and does what you have told it to. It can tell someone, wait for an approver, or act straight away by pulling the agent back, tightening the gateway or reopening the risk assessment. It only ever does things a person on your team could do in Aicura, and every step is on the record.

Say what counts as a problem

Each condition is one figure crossing a line for a set time, such as error rate at or above 20 per cent of requests for an hour. In the app the figure is called a measure and the condition a band. The figure is the same one monitoring shows, for the same system, in the same unit.

A condition can apply only to certain kinds of system or certain risk ratings, and can combine with other conditions, so a problem that is two facts at once is still one condition. Each one says whether it clears by itself when things recover, or stays open for a person.

Decide how far it can go on its own

Every condition sits at one of five levels, which the app calls rungs. The first two only watch and tell. The last three act.

Off
Not watched.
Tell me
Raise it with a person and stop there.
Act with approval
Wait for a named approver before doing anything.
Act and notify
Act, then say what it did.
Just do it
Act.

What it can do when something goes wrong

Every response Guardian can take already exists elsewhere in Aicura, ready for a person to use. Guardian only decides whether to use it, and how far from a person.

Tell people

  • Raise it in Guardian
  • Notify the owner
  • Notify the approvers for that area

Look again

  • Re-run the risk assessment
  • Re-run the impact assessment
  • Run a security test

Tighten live controls

  • Deploy the restricted guardrail set
  • Deploy the tightened budget set
  • Redact more personal information
  • Refuse requests that cannot be checked
  • Withdraw the deployed guardrails

Rein in the system

  • Narrow what the agent may call
  • Revoke the agent's credential
  • Refuse credentials it never declared
  • Put the system under review
  • Quarantine the system
  • Suspend a network component

Open follow-up work

  • Open an incident
  • Open a discovery item
  • Open a vendor review
  • Stop sampling allowed requests

You stay in charge of it

Set it for everything, or for one system

Choose what a setting applies to, from every system and agent down to one, and the most specific setting wins. Settings can also cover the components Aicura runs in your network, such as gateways. Each row says whether its level is Aicura's default or one you set, so you can always see why Guardian behaves as it does.

  • Every system and agent
  • By kind
  • By risk rating
  • One system or agent
  • The whole estate
  • By kind of component
  • One component

Pause the Guardian stops it acting at all, without changing any of your settings.

Approvals with a time limit

A condition that acts names its approver, whether a named person, any administrator or any operator, and how long to wait. While an item waits, it shows when the wait ends. When time runs out, the responses expire or go ahead, whichever you chose.

If a condition is set to act but is missing something it needs, such as an approver or the guardrail set to deploy, its settings say so and the dashboard counts it as blocked until it is fixed.

Everything it does is on the record

The Dashboard shows what is proposed, waiting for approval and already acted on, plus any conditions that are blocked. The History tab lists closed items with who decided and when. Each item shows the figure behind it, a plain line for each fact Guardian checked, and every response with its outcome and a link to what it produced, such as an incident or a vendor review.

Every watched AI system, agent and network component has its own Guardian tab. Items are snapped into the Evidence Vault, Aicura's store of dated snapshots, when they are raised and again when they close.

Questions

What is Guardian in Aicura?

Guardian is the part of Aicura that responds automatically when an AI system or agent runs into trouble in production. It watches the same figures monitoring shows, such as error rates or calls to tools an agent never declared. When a condition you set is met, it tells a person or takes an action a person could already take in Aicura, as far as you have allowed.

Does Guardian act without asking a person?

Only where you have told it to. Each condition Guardian watches is set to one of five levels. Off and Tell me only watch and report. Act with approval waits for a named approver, Act and notify acts and then reports, and Just do it acts. Guardian can be paused at any time without changing a setting.

Can Guardian do things a person on our team could not?

No. Every response Guardian can take already exists elsewhere in Aicura, such as deploying a guardrail set, narrowing what an agent may call or opening an incident. Guardian only decides whether to use it, and how far from a person.

What happens if a Guardian approver does not respond?

Each Guardian condition that acts has a hold time, a day by default, and you choose what happens when it runs out. The responses either expire unused or go ahead. Until then the item shows as held, with the time left.

How do we see what Guardian has done?

The Guardian Dashboard lists open items and the History tab lists closed ones, with who decided and when. Each item shows the figure behind it, the condition it met and every response with its outcome. Items are snapped into the Evidence Vault, Aicura's store of dated snapshots, when raised and again when closed.

Decide what your AI governance can do on its own

Contact us about the problems that should bring a person in, and the ones that should not wait for one.