Reviewing risks in a new technology proposal
Ask how a proposed technology changes data exposure, access, cryptography and recovery, then assign checks before rollout.

A new technology can create a different path to business information or a new dependency during recovery. Review those changes before the feature becomes part of ordinary work.
Keep the assessment tied to the proposal in front of you: an AI note-taker for client meetings, a fingerprint time clock, tablets on 5G for a field crew. A list of emerging threats doesn’t tell you which risks your business has.
What information enters the system?
For an AI tool, list what it will see: staff prompts, uploaded files, connected mailboxes or document libraries, and any reference material. Record who can change each source.
Then ask the supplier:
- Is our data used to train your models? Microsoft, for example, states that prompts, responses and data accessed through Microsoft Graph aren’t used to train the foundation models behind Copilot. Get the same in writing from any other supplier.
- What can it reach? Copilot only shows content a user already has permission to view. A folder shared with the whole company years ago becomes easy to find through it, so tidy permissions before you switch it on.
- Where is the data processed and kept, and can we delete it?
- How does it handle malicious or misleading input, such as instructions hidden in a document or email it’s asked to summarize?
The Cyber Centre’s AI security actions address data integrity and controls around AI use. Start with a limited trial on non-sensitive material.
A generated answer should remain subject to review even when the input looks familiar.
Which identities and devices gain access?
For a phone, tablet or connected-device proposal, ask who owns the account, how long the device will get security updates and who has administrator access. Ask whether it can be enrolled in device management such as Microsoft Intune, which can apply settings, wipe a lost device and remove only the work data from a personal phone when someone leaves.
Using 5G doesn’t by itself make a device unsafe or safely isolated. Check the connection itself, the device configuration and the path to business applications.
For headsets and smart glasses, include cameras, microphones and recordings of the room in the data-flow review.
How is sensitive identity information handled?
If biometrics are involved, establish whether matching happens on the device or in a supplier’s service, what is kept and who can access it. Unlike a password, a fingerprint can’t be changed if it’s exposed.
Implementations differ. Windows Hello stores its biometric data encrypted on the device and doesn’t send it to Microsoft. A time clock or door system may keep templates in the supplier’s cloud. Ask where templates are stored, whether they’re encrypted, how they’re deleted when someone leaves and what staff who decline can use instead.
The OPC’s business biometric guidance addresses purpose, necessity and safeguards. Have the privacy owner review the proposed use and alternatives.
What is the supplier’s cryptography plan?
Post-quantum planning is about a future quantum computer that could break the public-key cryptography in use today. Data copied now could be read then, which is why long-lived confidential records matter. The Cyber Centre’s guidance for organizations says such computers could be available as early as the 2030s. Current encryption isn’t broken, so don’t install unsupported replacements.
For a small business this is mostly a set of questions for suppliers:
- When will you support the post-quantum algorithms NIST standardized in 2024 (FIPS 203, 204 and 205)?
- Will that arrive as an ordinary update, or will it need new hardware or a different plan?
- Is the cryptography you use now standardized and validated?
Note which records must stay confidential for many years, such as client files or contracts, and bring those suppliers’ answers into lifecycle planning.
What happens when the feature fails?
Test the fallback, record export and recovery process. Include the possibility of stolen information as well as unavailable systems.
AI also makes impersonation easier. The Cyber Centre’s social engineering guidance notes that a voice can be cloned from small audio samples, and that audio or video alone shouldn’t be taken as proof of identity. Decide now how staff confirm an unusual payment or data request: a call back to a number you already have, or a second channel agreed in advance.
Record required controls, unresolved questions and the person who accepts the remaining risk. To give each finding an owner and a deadline, see turning exposure findings into security fixes.


