Skip to content

How it stays safe

Agents do the work. Approval rules control sensitive actions.

On the controlled action path, actions are sorted by risk. Reading, drafting and reversible changes may run under their rules. Actions requiring approval wait for a person.

Ten requests, instant verdicts, no sign-up.

Try it

Ten requests. See what an agent does with each.

Ordinary work runs. Risky work waits for a person. The four attacks in the last group are refused.

Pick a request

The four under Try to break it are the ones worth your time. They are the attacks people actually worry about.

Ordinary work · 3

Risky work · 3

Try to break it · 4

1 of 10 tried

What happens

T3held

Held for a named person

Money leaving the business is T3. The agent prepared the whole refund (amount, reason, evidence) and stopped. A person approves it, and the record shows who.

The rule underneath

The agent does the work. A person makes the decision that cannot be undone.

record · T3 · refund.issue · held · recorded

This is a simulation of the rules, not a live workspace. On a call we run the same requests against a real one and show you the events as they are written.

Why it comes out that way

Four rules behind every answer.

  • A person approves what cannot be undone

    Payments, deletions and deployments wait for a named person. A model’s “yes”, typed or spoken, never counts.

  • Agents only see the tools they are allowed

    Tools are filtered by policy before the model sees them, so a tool off the list cannot be called.

  • Everything is logged, refusals included

    Governed actions leave a record, whether they finish, fail, are refused or await approval.

  • Important results face an independent challenge

    On supported delivery runs, a separate model family tries to find defects before work can close. Material findings return the work for correction.

Risk tiers

Governed actions, sorted by risk.

This is the whole table. You can adjust the defaults, and changing a rule is itself an action only designated owners can take.

  • Risk tier
    T0
    What it covers
    Read
    Example actions
    Search, read, summarise
    Who may execute
    Any agent with a resolved policy
    Recorded
    Request and result
  • Risk tier
    T1
    What it covers
    Draft
    Example actions
    Draft, prepare, propose
    Who may execute
    Any agent with a resolved policy
    Recorded
    The draft and its sources
  • Risk tier
    T2
    What it covers
    Reversible change
    Example actions
    Update a record, schedule, tag, assign
    Who may execute
    Agents, up to their ceiling
    Recorded
    Before and after
  • Risk tier
    T3
    What it covers
    External or hard to reverse
    Example actions
    Send to a customer, publish, spend
    Who may execute
    Approved wording may send under policy; other external actions need a named person
    Recorded
    Approver, evidence, policy version
  • Risk tier
    T4
    What it covers
    Irreversible or regulated
    Example actions
    Pay, delete, deploy, submit a filing
    Who may execute
    Person only, agent assists
    Recorded
    Approver, second approver where set, full evidence
  • Risk tier
    T5
    What it covers
    Change the rules
    Example actions
    Edit policy, raise a ceiling, add a capability
    Who may execute
    Designated owners only
    Recorded
    Change, reviewer, blast-radius preview

Threats and controls

Six ways agents go wrong, and what stops each.

Each control is labelled Live or Partial. Detail on partial controls goes in the security pack, on request.

  • An agent does something destructive

    Live

    On the governed action path, actions are sorted by risk before they run. Actions requiring approval wait for a person.

  • An agent approves its own work

    Live

    Approvals accept a human only. A model’s confirmation, typed or spoken, never satisfies one.

  • A web page or email tells the agent what to do

    Live

    Retrieved content is treated as data, not instructions. The model also never holds the authority an injection is after, so fooling it gains nothing.

  • The model is confidently wrong

    Partial

    On supported runs, a different model family checks the work and a goal closes only with evidence and two approvals. Today that is software delivery. Business work is being productised.

  • A loop runs up a bill

    Partial

    Each run has step and cost limits, loop detection and circuit breakers on model providers. Workspace-wide spending ceilings are in preview.

  • One customer’s data reaches another’s

    Partial

    The workspace boundary is enforced in the database, not only in application code. Coverage is still being extended and independently tested.

A downloadable evidence export and single sign-on are not built. The security page lists every control with its status.

Questions

What people ask next.

A prompt guardrail is a request to the model, and a model can be talked out of a request. These checks run outside the model, between the agent’s intention and the action. The model does not get a vote.

Bring us the job you would never hand to an agent.

We will show you which parts an agent can do alone, which need a person, and which should stay fully human.

See a proposed package before sharing contact details. A person then reviews your answers and sends a scope and figure.