This is general educational guidance, not an assessment of your business or a guarantee of security. Check the original sources and your specific requirements before acting.
How it can happen
Imagine a business assistant reading a supplier document. Hidden or misleading text inside that document tells the assistant to send files somewhere else or change a payment destination. That text is part of the document, not a legitimate instruction from the business owner.
Design to limit the damage
There is no single filter that makes prompt injection impossible. Stronger systems control the actions an assistant is allowed to take, regardless of what it reads.
- Keep external content separate from authorised instructions.
- Grant tools and data access only when needed.
- Require meaningful human review before financial changes, outgoing emails or record changes.
- Check generated outputs before they enter business systems.
Reduce exposure
Avoid pasting passwords, confidential files or customer details into tools without an appropriate business reason and an understanding of the provider’s data handling. Review AI provider permissions and logs.
Read the original guidance
The links below are from the original publishers. This article is independently written by TASII and does not imply endorsement.
- Using AI securely in your businessASD’s Australian Cyber Security Centre · 7 October 2026
- Guidelines for secure AI system developmentASD’s Australian Cyber Security Centre and partners · 27 November 2023
- Defending against AI-enabled cyber attacks: Guidance for small businessesASD’s Australian Cyber Security Centre · July 2026
