Register for the
CSA Australia Summit

Agentic AI is already running on developer endpoints, in CI/CD pipelines and against identity infrastructure. The activity it generates lands in the same telemetry a SOC already collects, but it arrives stripped of the things analysts normally rely on to work out who did what. There is no reliable parent process, no consistent user context, and no signal in the raw event that separates a developer typing a command from an agent deciding to run one. Security operations teams are absorbing the consequences without a way to attribute activity back to an agent, which means AI forensics is largely guesswork right now. This session will walk through why OS-level evidence falls short, then set out a repeatable way to tie process execution back to the agent that caused it.

We start with the threats aimed at agents: prompt injection through untrusted content, poisoned skills and configuration files, malicious extensions and connectors, and abuse of the long-lived credentials agents hold. From there we trace what happens mechanically when an agent is asked to do something routine on a device, from tool invocation down to shell execution, and what that leaves behind in endpoint, cloud and identity telemetry.

The immediate problem for a SOC is volume. Agents drive nearly everything through the command line, so behavioural detections that used to be low noise and high fidelity (e.g. secret access, credential file reads, curl-based ingress and egress) now fire constantly. Filtering AI-initiated activity wholesale isn't an option, because a manipulated or compromised agent produces the same primitives as a legitimate one. Alerting on every agent action isn't workable either.

Most of the session is on attribution. Agents typically spawn an interactive shell and execute from inside it, so the agent binary never shows up as the parent process and conventional lineage checks miss it. We walk through using graph semantic operators in KQL to reconstruct full process ancestry and attribute arbitrary execution back to an agent, then treat that attribution as a triage dimension rather than an off switch. The last part is implementation: separating user-initiated from AI-initiated activity, layering malicious signals over the top, and correlating alerts into agent-scoped incidents.

Media manipulation involves deliberately manipulating media materials using AI-driven tools. The unprecedented proliferation of media manipulation on social media platforms can significantly influence public opinion and adversely affect society. AI-enabled tracking of the entire media manipulation pipeline can effectively combat this issue by providing authentication and provenance services for non-expert users.

This talk will provide key insights into the various factors that contribute to media manipulation, from the sources of media to the scenarios of manipulation and the complexities inherent in the process. We will discuss a tool currently in development that can assist in authenticating the origin of media and its manipulation as it propagates across social media platforms, and provide evidence-based justification and explanations for non-expert users.

For organisations using one or more of the cloud hyperscalers, obtaining visibility and monitoring of IaaS, PaaS and SaaS environments is difficult — especially at a velocity that keeps pace with the hyperscalers' offerings, the business's ambitions, technology deployment and solution architecture. Matching that to constantly evolving threats, triaging impacts and presenting the context is even more difficult.

Cloud Security Posture Management (CSPM) tools are a category of tooling to assist in this journey. This talk is about getting better at understanding the security posture of cloud-based delivery of business and technology services: implementing, learning about and driving adoption of different CSPM tooling. The tooling choice is complex and, as with any new technology stack, comes with complications that have very little to do with the tooling itself. The journey is not yet finished, and there will be reflections on what worked, what could be done better, and some attempt at guidance.

This presentation explores the concept of Human-in-the-Loop (HITL) Cyber Defence, where artificial intelligence and human analysts collaborate to strengthen organisational security. Drawing on research in traffic anomaly detection and cyber-attack detection, the session examines how AI can continuously monitor network activity, identify suspicious behaviour, and prioritise potential threats at scale.

Human analysts, in turn, provide critical judgement, contextual awareness, and oversight to validate findings, interpret complex situations, and make informed decisions in high-risk scenarios. By combining the speed and scalability of AI with human expertise, organisations can improve threat detection, reduce false positives, and build more resilient and trustworthy cyber defence capabilities.

Enterprise AI is easy to demonstrate. Putting it securely into production is much harder. As organisations move from GenAI experimentation toward AI embedded in real business processes, the challenge quickly shifts from the model itself to the enterprise environment around it: fragmented data, legacy applications, identity and access, cloud architecture, integration, governance and security.

This session explores the practical gap between an impressive AI pilot and a secure, scalable enterprise capability. Drawing on experience delivering complex digital transformation, data, cloud and AI initiatives, it examines where AI programs commonly encounter friction, why security and governance need to be designed into the architecture rather than added later, and what technology leaders should consider before allowing AI to interact with enterprise systems, data and workflows.

The session will provide a practical framework for moving from AI experiment → secure production → enterprise scale, focusing on the foundations organisations need to establish now as AI becomes increasingly embedded in their technology environments.

An introduction to BTCScamFinder, an open-source forensic pipeline that links scam website detection to Bitcoin transaction tracing — built and evaluated as a solo research project using only public data.