Your agent, tested on your terms.
Your chosen environment, your key, your limits. We test an agent only after its owner has confirmed the environment and the goals in writing.
What we touch
-
The environment you choose
We test against the environment you choose, ideally staging.
-
Nothing installed
We call the endpoint your agent already answers on.
-
Your agent’s actions
If it can take actions, such as payments or account changes, make sure they can’t reach real customers or live systems.
What we hold
-
Your documents
Read only for your assessment, by us and by the models agreed with you in writing. We don’t train models on your data.
-
The conversations
Every transcript, kept with your report.
-
No customer data
We don’t need chat logs. We ask for test or public data only. If the environment you choose holds real personal data, tell us before we start, so we can agree what to do.
-
Your material stays yours
Your documents, conversations, and findings are never used in another client’s assessment. We reuse only our own testing methods, written to carry no client detail.
What you control
-
Your key
Issued by you, carried on every request, revocable at any time. Revoke it and testing stops.
-
Your limits
We test inside limits you agree with us.
-
Your reports
Sent only to the people you name.
Agreed in writing before we start
- Your chosen environment and your goals
- Every model we use, and its provider
- Who receives your reports
- How long we keep our copy, in line with your policies
Findings built for scrutiny
-
A separate judge
The judge is a model from a different provider to every model that talks to your agent. It never sees your threshold or other conversations, so it rules on what was said.
-
The exact words
Every failing ruling must quote your agent. Code, not a model, matches each quote word for word. A ruling we can’t back with a quote never counts.
-
The agreed goal
The judge rules only on the goals you confirmed. If it isn’t in the goal, it isn’t a finding. The report can still note it.
Where the rules stand today
As of October 2026
-
Transparency
Since 2 August 2026, the EU AI Act has required AI that talks to people to be designed so they know it’s an AI, unless that’s obvious.
-
High-risk uses
From 2 December 2027, AI used for purposes such as assessing creditworthiness must be tested ‘against prior defined metrics and probabilistic thresholds’ before release or first use.
-
UK financial services
The FCA has said: ‘We do not plan to introduce extra regulations for AI.’ It will rely on existing frameworks, including the Consumer Duty.
-
A tribunal decision
A tribunal ordered Air Canada to pay for its chatbot’s wrong answer about bereavement fares.
A Mantia report is evidence for your own release decision. This section is general information, not legal advice.
Running a security review? Ask us through our contact form.
Cars are crash-tested to find their weak points. AI agents need the same.
We push yours to its limits, so you know when it’s ready.