For your board
When AI starts making or shaping decisions, someone independent should check it.
A briefing for directors and senior leaders: why independent testing matters, the language of AI risk in plain English, and the questions to ask your teams.
Download the board briefing (PDF)Start here
Five questions every leader should be able to answer.
What AI do we use, including AI built into tools we already pay for?
What can each system access, and what is it allowed to do?
When was it last tested, and by whom?
Would we know if it was manipulated or started giving wrong answers?
What evidence could we show a regulator, client or insurer tomorrow?
If any of these are hard to answer, that's the gap independent testing closes.
Why independence matters
If the same firm helped design or implement your AI, it shouldn't also be the one that independently checks it. Keep your existing advisers for strategy and delivery, and use FireBreak to check the result.
Why it takes specialists
Testing AI needs people who know how it fails in practice: through manipulation, confident wrong answers, unfair outcomes and agents doing more than intended.
Why AI risk is different from ordinary IT risk.
It's learned, not written
Nobody writes an AI's rules, so nobody can read them. You can't inspect your way to safety; you have to test how it behaves.
Instructions and information arrive as the same text
A line hidden in an email or document can be read as an order. Anyone who can write a sentence can attempt an attack.
It changes after launch
Its behaviour can shift when the model is updated, new data is added or it's given new tools. A pass last month isn't a pass this month.
It can act, not just answer
AI agents send emails and update records. The risk is a wrong action, taken at speed, with whatever access the agent has.
The language of AI risk
The terms your teams use, and the question to ask about each.
Prompt injection
Tricking the AI with words
Someone hides instructions in a message, email, document or web page, and the AI follows them instead of yours.
Ask your teamCould a stranger give our AI instructions, and would we know if they did?
Jailbreaking
Talking the AI out of its rules
Someone persuades the AI to ignore its safeguards and say or do things it was told not to.
Ask your teamHave our safeguards been tested by someone trying hard to break them?
Data leakage
The AI says too much
The AI reveals what it shouldn't: customer records, internal documents, or the hidden instructions it runs on.
Ask your teamWhat is the most sensitive thing our AI can see, and could it be coaxed into sharing it?
Data poisoning
Feeding the AI false information
Someone plants misleading content in the documents or data the AI relies on, so its answers quietly become wrong.
Ask your teamWho can change the information our AI relies on, and is that controlled?
Hallucination
Confident wrong answers
The AI makes something up and presents it as fact. Courts have held organisations liable for what their chatbots promised.
Ask your teamWhen our AI is wrong, who is accountable, and would we notice?
Rogue agents
AI doing more than intended
An AI agent that can take actions, such as sending emails, changing records or making payments, goes beyond what it was meant to do.
Ask your teamWhat is the worst thing our AI is allowed to do without a person approving it?
Unfair outcomes
Treating people differently
The AI gives worse answers or decisions to people because of their name, age, accent, disability or background.
Ask your teamHave we tested whether our AI treats different groups of people equally?
Model drift
The AI changes after launch
When the supplier updates the underlying model, the AI's behaviour can change without anyone at your organisation changing anything.
Ask your teamWhen did we last test our AI, and has the model changed since?
Free guide
When AI Fails: what every leader needs to know
Our guide for leaders: what the 2026 research shows, why AI risk is different, how AI fails in practice, with real incidents, and what to ask your teams.
A low-risk first step
AI Risk Scorecard
One AI system, tested across security, reliability, fairness and governance, and mapped to the EU AI Act. A board-ready report in two to four weeks, at a fixed price.
See the scorecard and a sample report- Report in
- Two to four weeks
- Your team's time
- About 4 hours
- Price
- Fixed, agreed before we start
Who we work with
- Scottish Government
- NHS 24
- Thoughtworks
Led by Jora Gill, former Chief Digital Officer at The Economist and SHL and CTO at S&P and RELX, who sits on FTSE boards.
The first external company on the Scottish Government's AI Register