Tools & Resources / AI assurance
Is Your AI Secure? Try Our Self-Assessment Framework
Artificial intelligence is no longer confined to research teams and experimental pilots. It now sits inside everyday business processes — embedded in products, delivered through third-party platforms, and updated continuously in the background.
AI assurance starts with the data.
As AI adoption has accelerated, so too has the complexity of managing the data risks that come with it. Many organisations are discovering that controls designed for traditional systems — or even early AI projects — no longer provide adequate assurance.
Interactive / Self-assessment
Take the AI System Data Security Self-Assessment
Understand where your AI risks are managed — and where they need attention.
Seven connected domains. Plain-language questions.
Internal AI · Third-party services · Embedded AI
01 / Data exposure
Why AI data security is no longer just about training data
Early conversations about AI risk tended to focus on a narrow set of concerns: where training data came from, whether it was biased or poisoned, and how models were initially deployed.
Those questions still matter — but they now represent only part of the picture.
Modern AI systems introduce additional data risks through:
- user prompts and uploaded content
- AI-generated outputs that may contain inferred or sensitive information
- operational logs and telemetry
- fine-tuning and continuous learning
- reliance on externally provided or embedded AI services
In practice, this means that AI data risk is now dynamic, distributed, and often outside direct organisational control.
Research suggests that more than 80% of employees now use AI tools at work — many without formal approval. Meanwhile, only a small fraction of organisations have developed mature AI security strategies. The gap between adoption and assurance continues to widen.
02 / The assessment map
What this assessment actually examines
The updated AI System Data Security Self-Assessment looks at AI risk across seven connected domains, covering the full lifecycle of how AI systems are used in real organisations:
01
Systems & ownership
knowing which AI systems you rely on and who is accountable for them
02
Data flows
understanding how data enters, moves through, and exits AI systems
03
Model behaviour
assessing how models behave, drift, and fail over time
04
Dependencies
managing the security and resilience of AI dependencies
05
Governance
ensuring meaningful governance and oversight
06
Change
monitoring change rather than assuming stability
07
Incident readiness
preparing for AI-specific incidents, not just traditional breaches
Importantly, the assessment applies regardless of how AI is delivered — whether built internally, consumed as a service, or embedded in third-party products.
03 / Third-party AI
Why third-party AI deserves special attention
One of the most common failure points we see is the assumption that AI risk can be outsourced.
In reality:
- accountability for data protection and outcomes remains with the organisation
- vendor AI can change behaviour without notice
- visibility into data handling and model updates is often limited
- contracts and assurances do not always reflect operational reality
The updated assessment therefore explicitly highlights third-party critical areas — questions where a lack of control or assurance in vendor-provided AI creates materially higher risk.
A “No” in these areas does not mean failure. It means attention is required, particularly before AI use is expanded or relied upon more heavily.
A practical example: In 2024, Air Canada was held liable for incorrect information provided by its website chatbot. The court rejected the argument that the AI operated independently. Organisations remain accountable for what their AI systems — including those provided by third parties — tell customers and stakeholders.
04 / Governance context
The regulatory context is shifting
Beyond immediate operational risks, the regulatory landscape for AI is maturing rapidly. The EU AI Act is now in force, with key obligations rolling out through 2025–2027. DORA introduces additional requirements for financial services. And enforcement actions — including fines for misleading AI claims — have already begun.
Organisations that build AI governance foundations now will be better positioned to demonstrate compliance as requirements become more specific. Those that wait may find themselves retrofitting controls under time pressure.
05 / Read the signals
How to interpret your results
The assessment uses simple Yes / Partly / No responses, but the results are deliberately explained in plain language.
Yes
A control you can support with evidence.
Partly
A useful signal: practice or evidence needs attention.
No
An area to examine and prioritise.
In our experience:
- “Partly” is the most common and most useful answer
- it typically indicates informal practices, inconsistent application, or limited evidence
- these gaps are often addressable without major technology investment
The goal is not perfection. It is clarity — knowing where risk is understood, where it is assumed, and where it is unmanaged.
A tool for governance, not just compliance
This assessment is designed to support:
- informed conversations between technical and non-technical stakeholders
- clearer accountability for AI systems
- evidence-based decision-making as AI use evolves
Used periodically, it provides a practical way to track progress and adapt controls as AI systems — and their risks — change over time.
AI can deliver real value. But only when its data foundations, dependencies, and behaviours are properly understood.