Investigating cyber threats faster while keeping humans in control

Australian organisations in critical sectors such as finance, government and health are under growing risk of cyber attack. The early warning signs of an incursion are often already sitting in the network data these organisations collect – the challenge is making sense of it quickly enough to prevent a breach.

A University of Queensland (UQ) team, working in partnership with Cisco, has built FlowMind: a secure agentic AI platform that turns raw network traffic into clear evidence on which an analyst can act in real time. The agent is restricted to a controlled, auditable workflow to protect the sensitive data these organisations hold.

What's the problem?

Cyber risk is rising across every sector Australians rely on – government, finance, health, telecommunications and education. In FY2024–25, the Australian Signals Directorate received more than 84,700 cybercrime reports, roughly one every six minutes.

Many organisations already have the data that could catch these attacks early. Network flow data records how systems communicate and can reveal the first signs of an attack: scanning, probing, unusual access, failed connections or a possible compromise.

The challenge for security analysts is interpreting these data quickly enough to take action – a single record of communication between devices might be completely harmless, or it might be the first clue in a much larger attack. They must examine ambiguous activity, decide whether it is suspicious by comparing it to historical data, and document their reasoning, all under time pressure.

What's UQ's innovation?

FlowMind is a secure, agentic AI platform that helps analysts investigate suspicious activity more quickly and consistently. It was built by researchers from the School of Electrical Engineering and Computer Science, in collaboration with Cisco, a world leader in networking technology.

FlowMind undertakes targeted evidence gathering to produce a report that includes a priority level and recommended next steps for a human to review. An orchestrator acts as a secure control layer that sits between the AI and the data, so AI only accesses the data it needs, and its actions remain governed and always documented.

How does the technology work?

An investigation can start from a single suspicious activity, an alert from another system, a scheduled threat-hunting task, or a manual request from an analyst. FlowMind analyses the data that it has, then works through a series of practical investigative questions. Is this activity common or rare? Did it contact many systems or just one? Does the pattern look like scanning, failed access, ordinary service behaviour, or something more serious?

These questions guide targeted evidence gathering: every query is tied to a clear purpose and runs only through the secure orchestration layer, never as a broad uncontrolled search.

Large language models, kept on a leash

The reasoning is done by a large language model – the same class of AI that powers conversational assistants, but here pointed at network evidence rather than open-ended chat. It reads what the orchestrator retrieves, proposes explanations, suggests the next question worth asking, and writes up its findings in plain language. A second review stage can re-run the reasoning to challenge the first assessment, a check that helps filter out false positives before anything reaches a person.

FlowMind is both agentic and secure. The model can choose which steps to take, but only inside a fixed harness. The orchestration layer decides which data the AI can see, which tools it can call, what actions are permitted, and how every step is logged. The model reasons, but it never reaches into the systems it’s investigating. That separation allows FlowMind to combine the powerful reasoning of large language models with the security of a governed investigation system.

Transformer neural networks for reading network traffic

FlowMind’s ability to make sense of raw telemetry rests on years of the team’s machine-learning research into flow data — most notably FlowTransformer, their framework for building intrusion-detection models out of transformer neural networks.

Transformers are the same architecture that underpins today’s large language models, but instead of learning the patterns in sequences of words, FlowTransformer learns the patterns in sequences of network flows. Reading traffic as a sequence, rather than one connection at a time, lets the models pick up longer-range behaviour — such as a drip of failed connections over time — that flow-by-flow methods tend to miss. In the team’s testing, the approach could match the accuracy of larger models at less than half the size.

Bringing years of research into one platform 

FlowMind draws together several strands of the UQ team’s prior work — including network-flow analysis, intrusion and anomaly detection, and human-in-the-loop security operations. Earlier research, including the FlowTransformer framework, established that network telemetry can expose suspicious behaviour, but that interpreting it at scale remains hard, and that language models need careful governance before they can be trusted in sensitive settings. FlowMind brings those ideas together in a single working platform.

By leveraging advanced AI models, we can create tools that detect and mitigate cyber-attacks in real time.

- Professor Marius Portmann, UQ-Cisco Chair of Network Security

What's the impact?

FlowMind’s most immediate value is making investigations faster, more consistent and more defensible. It can surface unusual patterns, help an analyst decide whether suspicious activity warrants escalation, assist with threat response, and produce reports.

The longer-term impact is stronger cyber resilience. As threats grow, organisations need tools that let them act earlier and more confidently, without lowering the governance standards that security-sensitive environments demand. FlowMind lightens the load on analysts while keeping a clear, auditable trail for every decision.

Its most important contribution to AI research and development lies in demonstrating how AI can be used responsibly: powerful reasoning kept inside secure boundaries, with transparent workflows and human oversight.

What's next?

Project details