May 25, 2026ZynexAI Security & Risk Management: How to Build Safe AI Systems
AI Security & Risk Management: How to Build Safe AI Systems
What Is an AI Security Assessment?
An AI security assessment is a review that tests the parts of an AI system a traditional penetration test doesn't touch: the prompts, the retrieved documents, the model's permissions and the vendor's data handling. A pen test looks for open ports and weak passwords. An AI review asks whether a cleverly worded customer email can make your assistant forward an invoice to the wrong person.
You still need the normal controls. The ACSC's Engaging with Artificial Intelligence guidance is clear that AI systems should sit inside your existing cyber security baseline. The AI layer just adds new ways in.
Eight AI Security Risks to Check Before Deployment
An AI risk assessment for an Australian business should check eight areas, from data flows through to staff use. Here's the list we work through, roughly in the order problems tend to show up.
1. Data Flows, Hosting and Retention
Every AI system moves data somewhere: to a model provider, a vector database, a logging tool. The first job is a map. Which systems feed the AI, which region hosts each one, how long is data kept, and is any of it used for training? We regularly find a test environment quietly sending real customer records to an offshore API.
2. Permissions and Access Control
An AI assistant connected to your CRM with an admin account can read everything that admin can. That's usually far more than the use case needs. We check service accounts, API scopes and whether the AI inherits the permissions of the user asking the question. Least privilege matters more here than almost anywhere else.
3. Direct and Indirect Prompt Injection
Prompt injection is an attack where text placed in front of the model overrides its instructions. Direct injection comes from the user typing. Indirect injection is sneakier: instructions hidden in an email, PDF or web page the AI reads. We test both with realistic payloads against your actual workflows, not generic jailbreak lists.
4. Leakage of Personal and Commercial Data
A support chatbot trained on your knowledge base can still repeat an internal pricing sheet or a staff phone number if those documents were indexed. We probe for leakage of personal information, commercial data and system prompts, and check output filtering before anything reaches a customer.
5. Vendor Terms and Data Handling
Your AI stack probably includes a model provider, an orchestration tool and a few plugins. Each one is a vendor with its own terms. The OAIC's guidance on commercially available AI products expects businesses to do due diligence on these tools, including how they handle personal information. We review contracts, data processing terms, training opt-outs and where each vendor stores data.
6. Automated Decisions and Privacy Transparency
If your AI helps decide who gets credit, a job interview, a price or eligibility for a service, that's an automated decision risk as well as a security one. From 10 December 2026, APP entities must explain certain substantially automated decisions in their privacy policies. The OAIC has published resources on this transparency requirement. We flag every decision point so your privacy team can check it. The full compliance audit is a separate piece of work.
7. Logging and Incident Response
When something goes wrong with an AI system, you need to know what it was asked, what it retrieved and what it did. Many SMB deployments log nothing useful. We check that prompts, tool calls and outputs are logged with personal data handled properly, and that your incident plan covers AI. If personal information leaks, the Notifiable Data Breaches scheme may apply.
8. Shadow AI and Staff Use
The AI you're formally deploying is rarely the only AI in the business. Staff paste contracts into free chatbots and connect browser extensions to email. We run a short survey and review of network and SaaS logs to find unapproved tools, then recommend a usage policy that people will actually follow.
What Does a Zynex AI Security Assessment Include?
An AI security assessment from Zynex includes a scoping workshop, technical testing and a ranked remediation plan.
How Long Does an Assessment Take and What Does It Cost?
An LLM security audit for a single chatbot or assistant is usually a few weeks of work, and cost depends mostly on how many systems the AI connects to. A standalone FAQ bot with no integrations is a small job. An AI agent that reads email, updates Xero and books jobs in your field service tool is much bigger, because every connection is a path for misuse.
Fixes are priced separately because they vary so much. Some are a config change. Others mean redesigning how an AI agent connects to your business systems.
What Zynex Sees in Australian AI Deployments
Zynex is a Melbourne-based technology consultancy that builds and secures AI systems for Australian SMBs, so we see these risks from both sides of the table. A few patterns come up again and again.
The model is rarely the problem. The plumbing is. The most serious findings in our reviews are usually an over-permissioned service account or a document store indexed without anyone checking what was in it.
Teams test the happy path. A chatbot gets tested with 50 sensible customer questions and passes. Nobody tries the angry customer who pastes in a fake "system message", or the supplier invoice with hidden white text.
Week three is where it slips. Once testing starts, owners realise nobody owns AI risk. We push for one named person before the assessment begins, which saves more time than any tool does.
Some projects aren't ready. If your data permissions are a mess across the business, fix that first. An AI assessment won't paper over it. Our AI Strategy Report is often the better starting point in that case.