Tech Translated

IT Security & Technology Blog

Practical IT insights for Australian businesses. Our team covers cybersecurity advisories, compliance updates, and plain-English explainers on the technology your business relies on, published regularly as the landscape shifts.

The AI Security Problem Australian Businesses Can't Patch Their Way Out Of

Australia's own cybersecurity agency has confirmed something worth sitting with: prompt injection attacks on AI systems cannot be reliably fixed inside the model itself. There is no patch coming for this one.

The Australian Signals Directorate (ASD) published guidance this week on the careful adoption of agentic AI services, concluding that prompt injection is architecturally inherent to how large language models work. The issue is that LLMs process user instructions and external content (documents, emails, web pages) inside the same context window. The model cannot reliably tell the two apart, so malicious instructions hidden inside a document or email can override what the actual user intended. ASD compared this to the early days of phone phreaking, where signaling and voice traveled the same channel, making certain attacks structurally possible regardless of how careful individual users were.

For Australian businesses, this is immediately relevant. Microsoft Copilot for Microsoft 365, ChatGPT Enterprise, and similar tools are now reading emails, files, and shared documents on behalf of staff. A successful prompt injection could cause an AI assistant to exfiltrate data, send messages the user never wrote, or take actions the user never authorised. ASD's recommended response is not to avoid AI tools, but to focus security controls on the "harness," the software layer surrounding the model: apply least-privilege access to what the AI can see and do, require human approval before the AI takes high-impact actions, log all AI interactions, and delete rather than summarise stale context from AI sessions. In practice, many Australian businesses are adopting Copilot through bundled Microsoft licensing via their IT provider, which means they often have limited visibility into what harness controls are in place and what data the AI is currently permitted to touch. That is the conversation worth having now.

Our cybersecurity team can review what AI tools are active in your environment and what controls are in place around them. It is a 30-minute conversation most businesses have not had yet.

Written by Caleb Attard, Technical Consultant, All IT Services. All IT is a Sydney-based managed IT provider supporting businesses across hospitality, not-for-profit, and professional services.


Frequently Asked Questions

Prompt injection is an attack where malicious instructions hidden in content (a document, email, or webpage) trick an AI assistant into following them instead of the user's legitimate instructions. Because language models cannot reliably tell the difference between instructions and content, there is currently no technical fix for this inside the model itself.
Yes. Any AI assistant that reads external content (emails, files, web pages) as part of its workflow is potentially vulnerable. That includes Microsoft Copilot for Microsoft 365, ChatGPT with browsing or file tools enabled, and most agentic AI products.
ASD recommends applying least-privilege access (limit what your AI tool can see and do), requiring human approval before AI takes high-impact actions, and logging all AI interactions. Ask your IT provider what controls exist in the harness around your AI product.
No. The right response is to deploy AI tools with proper governance, not to avoid them. The businesses most at risk are those who have rolled out Copilot or similar tools without thinking through what data the AI can access and what actions it can take.

Not Sure What Your AI Tools Can Access?

Most businesses using Copilot or ChatGPT have not reviewed what data the AI can reach or what actions it can take. We can help you map that out and put the right controls in place.


Related Guide

Cybersecurity for Sydney SMBs

Explore our complete guide to protecting your business from cyber threats.

Read the Full Guide →