AI Security Training
Your staff already use AI tools. The question is on whose terms.
What is AI security?
AI security training teaches staff which information must never go into an AI tool, how to recognise AI assisted fraud attempts, and gives them a sanctioned route so unapproved use stops being the only option.
Staff in most organisations are already pasting work into AI tools. Nobody sanctioned it and nobody forbade it clearly enough to stop it, so it happens quietly on personal accounts.
Banning it does not work. It moves the same behaviour onto phones where you cannot see it, which is worse than the problem you started with.
The two things that do work are a sanctioned tool people can actually use, and a short clear rule about what must never be pasted anywhere. Those change behaviour; a policy document does not.
Why a ban makes things worse
Duara runs AI security training from Nairobi for organisations handling client data, personal data or anything under a confidentiality obligation.
We tell clients that a ban without a sanctioned alternative makes things worse. It is not what a security team wants to hear, and the evidence in every organisation we have looked at is the same: prohibition moves the behaviour out of sight rather than stopping it.
Who this is for
- Staff are using AI tools on personal accounts
- You handle client or personal data
- A policy exists and nobody follows it
- AI assisted fraud attempts are increasing
Industries we serve
- Banking and financial services
- Legal and professional services
- Healthcare providers
- Insurance
- Public sector
- Any business handling client data
Problems this solves
These are behaviour problems rather than technical ones, which is why technical controls alone do not fix them.
- Shadow use on personal accounts. Work is pasted into consumer tools with no oversight, and the organisation cannot see any of it.
- A policy nobody reads. A document exists, was circulated once, and does not affect what anybody actually does.
- Client data leaving unnoticed. Confidential material is pasted in without anybody intending harm, because nobody made the rule concrete.
- AI assisted fraud getting better. Phishing and impersonation attempts are more convincing than the badly written examples staff were trained on.
What we deliver
A sanctioned tool with appropriate terms, one short rule about what must never be pasted, training on recognising better fraud attempts, and a route to ask when unsure.
A short rule staff can actually remember
A sanctioned tool with appropriate data terms
Training on recognising AI assisted fraud
A route to ask when somebody is unsure
Guidance for handling client and personal data
A realistic policy rather than an aspirational one
Features
- A sanctioned alternative first. A ban without one moves the behaviour onto personal phones, which is worse than the problem you started with.
- One rule, not a document. A short memorable rule changes behaviour where a policy nobody finished reading does not.
- Fraud examples that are actually convincing. Staff trained on badly written phishing are unprepared for messages that read properly.
- A route to ask. Somebody unsure should be able to check quickly, or they will simply guess.
- Policy written to be followed. An aspirational policy that everybody breaches is worse than a narrower one everybody keeps.
Benefits
- Shadow use becomes visible use. A sanctioned tool brings the behaviour back where it can be seen and governed.
- Client data stops leaking quietly. A concrete rule prevents the pasting nobody intended as a breach.
- Staff recognise better fraud. Training on convincing examples prepares people for what they will actually receive.
- Your policy becomes true. A rule that reflects real behaviour is enforceable in a way an aspirational one never is.
Technologies we use
- Data classification in practice. Making the distinction between what may and may not be shared concrete enough to apply quickly.
- Tool terms. Which arrangements keep your data out of training and which do not, since consumer and business terms differ.
- Fraud recognition. Identifying AI assisted phishing, impersonation and voice cloning attempts.
- Shadow use detection. Understanding where unsanctioned use is happening before deciding how to respond.
- Incident response. What to do when somebody realises they pasted something they should not have.
The security question people ask is whether AI tools are safe. The more useful question is which tool, on what terms, with what going into it, since those differ enormously.
How the process works, step by step
Security training is usually half a day, and the sanctioned tool question is settled before the session rather than raised during it.
-
Step 1
Free exposure review
We establish what your staff are probably already doing, which is usually more than management expects.
-
Step 2
Sanctioned tool decided
A usable approved option is settled first, because training without one produces guilt rather than change.
-
Step 3
The rule written
One short rule is agreed, deliberately narrow enough that people will actually keep it.
-
Step 4
Training delivered
Staff learn the rule, the reasoning and how to recognise convincing fraud attempts.
-
Step 5
A route to ask established
Somebody unsure has a quick way to check rather than guessing.
Quality assurance and how we work
The commitments attached to advising on how staff use these tools.
- We say a ban alone makes things worse. Prohibition without a sanctioned alternative moves the behaviour onto personal phones where you cannot see it.
- We write policies that can be kept. An aspirational policy everybody breaches is worse than a narrower one everybody follows.
- We use convincing fraud examples. Staff trained on badly written phishing are unprepared for messages that read properly, which is what they now receive.
- We do not oversell technical controls. This is largely a behaviour problem, and controls that do not account for that get worked around.
- We tell you what staff are already doing. The exposure review usually finds more use than management expected, and reporting that accurately is the point.
Frequently asked questions about AI security
Should we just ban AI tools?
No, and we will say so plainly. A ban without a sanctioned alternative moves the behaviour onto personal phones where you cannot see it.
Are our staff already using them?
Almost certainly, and the exposure review usually finds more use than management expected.
What is the single most important rule?
Nothing client confidential, personal or under a disclosure restriction goes into a tool you have not approved. Short enough to remember is the point.
Why not a full policy document?
Because a document nobody finished reading does not change behaviour. A short rule people keep is worth more than a thorough one they breach.
Are business AI tools safe to use?
Safer, in that business terms usually exclude your content from training. The question is which tool on what terms rather than whether AI is safe.
What counts as shadow AI here?
Unsanctioned use on personal accounts, which is the normal state in organisations that have not provided an approved option.
How is AI changing fraud attempts?
They read properly now. Staff trained to spot badly written phishing are unprepared for messages with no obvious errors.
What about voice cloning?
It is a real and growing risk, particularly for payment authorisation by phone, and it deserves a process change rather than only awareness.
What if somebody has already pasted client data?
There is an incident process and it should not begin with blame, or the next person will not report it.
How much staff time does this take?
Half a day, with the sanctioned tool question settled before the session rather than raised during it.
Why settle the tool first?
Because training without an approved alternative produces guilt rather than change. People need somewhere to go.
Does this cover prompt injection?
Yes, at a practical level for staff, and in more depth where your organisation builds AI systems rather than only using them.
Can you write our policy?
Yes, and we will write one that can actually be kept rather than one that reads well and gets breached weekly.
How do we tell whether behaviour changed?
Fewer people using personal accounts, and incidents being reported rather than hidden. The second is a better signal than the first.
Should staff be disciplined for shadow use?
Generally not, before an approved alternative exists. Punishing people for a gap you created damages reporting more than it changes behaviour.
Does this suit regulated sectors?
Yes, and banking, legal and healthcare clients usually need the tool terms question answered in more detail.
Can it be delivered to a whole organisation?
Yes, though smaller sessions allow the awkward questions that are usually the most useful ones.
Do you cover the Data Protection Act?
At a practical level. Your data protection officer or legal adviser confirms your specific obligations.
Will you report what individuals told you?
No. Sessions surface behaviour patterns, and we report those without naming people.
What does security training cost?
Group size and session count determine it, quoted after the exposure review.
Glossary of terms
AI security terms in plain language.
- Shadow AI
- Unsanctioned use of AI tools by staff, usually on personal accounts.
- Data classification
- Deciding which information may be shared where, made concrete enough to apply in seconds.
- Prompt injection
- Hidden instructions in content that manipulate an AI system reading it.
- Voice cloning
- Synthesised speech imitating a specific person, increasingly used in fraud attempts.
- Training data
- Material used to improve a model, which business terms usually exclude your content from.
- Incident
- An occasion where information went somewhere it should not have.
Related services in AI Training
Talk to us about AI security
Ask your team how many of them used an AI tool for work last week. The honest answer is usually higher than your policy assumes.
Start your project