Agentic AI Has Changed the Rules of Governance: Here's Where to Start

By Allison Lawrence, Senior Legal Writer, Practical Guidance Technology & Innovation, LexisNexis® Australia

A brief introduction to the new governance concepts every legal leader needs to know, and where to go to get up to speed.

This article is a brief introduction to some of the key concepts shaping the Australian regulatory guidance conversation about AI. The concepts are unpacked in detail inside our new whitepaper, ‘Governed to Grow: AI Governance as a Foundation for Innovation’, which also includes a practical checklist you can use immediately.

In June 2026, an AI agent doing nothing more sinister than researching public medical spending found its way into non-public files at Services Australia. It wasn't hijacked by a malicious actor. In pursuit of its task, the agent simply encountered a restriction – and, moving beyond the access boundaries intended for it – it found a way around it.

By the time the Prime Minister disclosed the incident in late September, the story wasn't only about the unauthorised access anymore. It was also about the almost three-month gap before anyone was told, and the pointed questions that followed about who was accountable, who was watching, and who should have known sooner.

For legal leaders, that incident is a useful marker. It shows how quickly "AI governance" has expanded beyond a compliance conversation about outputs — accuracy, bias, hallucination — to something much bigger: a live, structural risk running through contracts, procurement, cybersecurity, and operational resilience simultaneously. Clients are already living with this shift, whether or not their advisors are ready to talk about it.

Meet the "Harness": a New Centre of Gravity for AI Risk

Until recently, advising on AI meant advising on the model — its licence terms, its training data provenance, its output warranties. But that's no longer the whole risk picture.

Regulatory guidance from the Australian Signals Directorate has introduced a concept worth adding to every technology lawyer's vocabulary: the “agentic AI harness”. Think of it as the scaffolding around the model — the layer that decides what data the AI can see, what tools it can call, what permissions apply, and whether its proposed actions get executed. The model reasons and proposes. The harness helps determine what’s allowed to happen.

Why does this matter? Because it means the highest-impact risks may not live in the AI vendor's core product at all — they may emerge in the integration layer connecting it to your client's systems.

Understanding the harness reframes the questions worth asking about any AI deployment: who controls it, who's responsible for its permissions and connectors, who owns the audit trail, and who's accountable when an integration quietly changes what the AI is able to do.

During LexisNexis’ recent AI Summit, Greg Dickason, Chief Technology Officer at LexisNexis, argued that human oversight is the standard that AI cannot replace. Dickason describes legal AI in three connected layers: The AI model; the harness around it that grounds and checks the model and equips it with specialised legal tools; and the human system surrounding both.

This distinction is important, and one the whitepaper walks through in detail — including how the Australian Signals Directorate (ASD) suggests organisations use it as a durable control point that persists even as the underlying model changes.

This creates a new focus: harness and ecosystem governance. The section within the whitepaper on "What Lawyers Should be Advising Now" explicitly recommends that procurement and contracting should now cover the harness, connectors, permissions, logging capabilities, and responsibility allocation for configuring/maintaining controls — not just the underlying model licence. That's a concrete shift in tech/commercial practice beyond traditional AI vendor contracts.

Understanding Governance Reach: "How Much Do You Actually Control?"

The second concept is more conceptual, but no less important. Emerging Australian guidance is framing AI risk through the lens of “governance reach”, a deceptively simple question about how much of an AI deployment an organisation genuinely controls, versus how much sits with suppliers, partners, or other AI agents operating outside its boundary.

The whitepaper identifies three flavours of governance reach: arrangements where one organisation governs everything (singular), arrangements shared across multiple organisations under agreed terms (federated), and open environments where an AI system interacts with participants nobody fully controls. As reach decreases, reliance on contractual protections, monitoring, and assurance may need to increase to compensate.

For anyone advising clients on AI, this is a helpful mental model to have on hand: it turns a vague sense of "AI risk" into a structured question — where does our control stop, and what have we done to cover the gap?

The whitepaper sets out AISI's full framework for governance reach, along with the related idea of ecosystem risk — the recognition that a model can perform exactly as intended while risk still emerges elsewhere, through a supplier failure, a cloud outage, or an unexpected interaction between agents.

The whitepaper also offers a structured way to advise clients on where their contractual protections, assurance mechanisms, and monitoring obligations need to tighten as control decreases — a natural entry point for cybersecurity, procurement, and cross-border/supplier engagements.

Why Law Firms and Organisations Need to Get Up to Speed Now

The governance of AI systems, including frontier and agentic AI, is increasingly a board and executive responsibility. AI is no longer viewed solely as a technology issue. Increasingly, it is recognised as a source of governance, operational, cyber-security, resilience, privacy and compliance risk that requires active oversight by boards and senior leadership.

None of this is speculative. Australian government agencies and regulators — AISI, ASD, APRA, ASIC and OAIC among them — are actively building out this guidance in real time, and the Government has now launched a rapid review into whether existing legislative, governance and information-sharing arrangements are fit for purpose for AI-related cyber incidents. The rogue OpenAI agent incident impacting Medicare shows exactly how fast theory becomes headline. Clients are going to be navigating these questions with or without informed legal advice.

The sooner these concepts feel familiar, the sooner you're able to have a genuinely useful conversation when a client or stakeholder raises them.

And that’s ultimately the idea behind the Governed to Grow whitepaper. Good AI governance isn’t about putting the brakes on innovation. It’s about building the confidence and control needed to make innovation possible at scale.

Allison Lawrence is a Digital, Technology and Innovation lawyer and Certified Information Privacy Manager (CIPM) with extensive experience advising on technology transactions, commercialisation, data governance, AI compliance, and digital transformation and risk. Allison has worked across both private practice and in-house corporate roles, bringing a practical and commercial perspective to the governance of emerging technologies and the challenges of innovation at scale. Allison is currently the Senior Legal Writer at LexisNexis for Practical Guidance Technology & Innovation.

Governed to Grow: AI Governance as a Foundation for Innovation unpacks the harness, governance reach, ecosystem risk, and more in detail, including what's now expected of boards, legal teams, and technology leaders.

Download the whitepaper here →

Practical Guidance: Technology & Innovation. For ongoing analysis, frameworks and tools as this area continues to move quickly, explore the Technology & Innovation module here.