CSA-IMDA Advisory on Safe Generative AI Use: Rules for Staff

What the CSA-IMDA advisory is

CSA lists it as advisory AD-2026-008, titled Safe and Secure Use of Generative AI for Individuals, dated 28 July 2026. It is a short two-page PDF. It does not create new legal duties. It is guidance, in the same family as CSA's other public advisories.

It sits next to the PDPC's Advisory Guidelines on Use of Personal Data in Generative AI, published a week earlier. The PDPC guidelines speak to organisations and the PDPA. The CSA-IMDA advisory speaks to the person typing the prompt. A sensible workplace policy needs both: our summary of the PDPC generative AI guidelines covers the organisation side.

Rule 1: what staff must not put into AI tools

The advisory is blunt: once you share information with a generative AI tool, you lose control over how it is stored, processed or accessed. It lists full name, address and NRIC number, and sensitive information such as financial or medical records, as things to keep out. It adds that workplace information and confidential business data should never be given to AI tools without proper authorisation.

In practice, the risky pastes in a small business are predictable. Give staff a concrete list rather than a vague warning:

• customer names with phone numbers, emails, addresses or NRIC numbers;
• staff records, payslips, medical certificates or performance notes;
• bank statements, invoices with account details, and payment screenshots;
• contracts, tender documents and pricing you have not published;
• passwords, API keys, one-time codes and system screenshots that show them;
• anything a client has shared with you under a confidentiality agreement.

Authorisation is the key word. A company-approved business plan with training switched off and admin controls is a different thing from a personal account on a free tier. Our ChatGPT at work policy guide explains how to decide which tools are approved and what data each may see.

Rule 2: approved tools only, and watch the extensions

The advisory tells users to pick reputable tools with clear privacy policies and terms, and otherwise to assume the provider can access everything and may use conversations for model training. It specifically calls out third-party apps, such as unofficial mobile apps and browser extensions, which may collect and use your data.

This is the gap most offices miss. A staff member installs an AI writing helper in Chrome, and it can now read every web page they open, including the company webmail and accounting system. The advisory's red flags are useful as a staff test: permission requests that do not match what the app does, such as camera or location access, and apps from unknown developers.

Rule 3: verify AI output before it leaves the company

The second practice in the advisory deals with accuracy. It warns that guardrails are not foolproof and that AI tools produce hallucinations: fictional or partly correct information presented convincingly, including fabricated references, quotes and statistics. Its advice is to cross-check with official and trusted sources, consult professionals on health and legal matters, and treat AI output as a starting point.

For a business, wrong output is not just embarrassing. A made-up regulation in a client proposal, an invented price in a quotation or a wrong date in a customer email creates real liability. Set a simple sign-off routine:

1. 1

   Mark AI-assisted drafts

   Staff note in the file or email draft that AI helped, so the reviewer knows what to check.

2. 2

   Check every number, date and name

   Each figure must trace back to your own records or an official source, such as a gov.sg page, not to the AI's answer.

3. 3

   Open every link and citation

   AI tools can cite pages that do not exist or that say something else. If the source does not support the claim, remove the claim.

4. 4

   Get a professional view on legal, tax and medical points

   The advisory says this directly for high-impact issues. AI can help you prepare questions, not replace the adviser.

5. 5

   Pause before sending or posting

   The advisory's 'pause before sharing' step: does it make sense, and could you defend it to the customer?

Rule 4: people decide, and the account stays secure

The third practice is about healthy boundaries. For individuals it focuses on emotional dependence on AI companions and keeping critical thinking. The workplace version is simpler: AI can draft, summarise and suggest, but a person makes decisions about customers, hiring, refunds and money.

Account security belongs here too. An AI account holds months of prompts that can reveal a lot about your business, and the advisory notes that if a tool is compromised, that history can be misused for fraud or scams. Treat AI logins like email logins: company accounts, MFA switched on, and access removed when staff leave.

Scammers also use generative AI to produce convincing messages, voices and videos. The advisory does not go into deepfakes, but the same verify-first habit applies: a voice note or video call asking for an urgent payment change should be confirmed through a known phone number before anyone acts. Our phishing spotting guide covers the written version of this trick.

One-page staff checklist based on the advisory

Print this, or paste it into your onboarding pack. It covers the three practices in a form staff can follow without reading the advisory.

Ask staff to acknowledge it once and again each year. That record also helps show the PDPC that you have reasonable arrangements in place.

Do it yourself or get help?

A small team can adopt these rules in an afternoon: pick the approved tools, switch on business-plan privacy settings, circulate the checklist and get acknowledgements. The harder part is keeping the habit alive as new tools appear.

Our Security Awareness Training (S$390) sets up short monthly lessons in your own account, including Singapore scam topics and safe AI use, with completion tracking. Staff who know why the rules exist follow them better than staff who were just sent a PDF.