1. The Medicare incident is unlikely to be the last case of an AI agent accessing systems or data in ways that weren’t intended.What does that pattern tell you about the specific failure in how most organisations are deploying agents?
The specific failure is that most organisations are racing to deploy AI agents on an access model designed for people.
Organisations used to focus only on authenticating a user at the front door and then trusting that user to behave appropriately once inside. That assumption works reasonably well for humans. We move slowly, understand context, and intuitively stop when we encounter information we shouldn’t access.
Agents don’t operate that way. When they’re given a task, they’ll use whatever permissions are available to complete it. Our research found that 99.6% of Australian IT leaders experienced an AI tool or agent accessing sensitive data beyond its intended scope over the past year.
In most of these cases, the agent isn’t acting maliciously. It’s simply following instructions with the access it has been given. The real issue is that organisations are unintentionally handing agents excessive, inherited, persistent or unmanaged access, creating opportunities for them to reach data or systems that were never part of the original task.
2. The agent was just looking for publicly available information” is the kind of explanation that obscures more than it reveals. What needs to change about how organisations define and enforce that boundary?
Organisations need to change how they define and enforce boundaries around access.
The Deputy Prime Minister described the incident as data sitting behind a fence that the agent climbed over. Regardless of the specific technical circumstances, the broader lesson is that organisations can no longer rely on boundaries that assume users will naturally stay within expected limits. Humans generally understand context and intent. Agents pursue objectives and will follow paths available to them unless controls explicitly prevent them from doing so.
The boundary therefore needs to be defined by what an identity is permitted to access, for what purpose, for how long, and under what conditions.
Most importantly, those controls can’t rely on the agent regulating itself. Enforcement needs to sit outside the agent, continuously validating access requests in real time and denying actions that fall outside approved policy.
3. Excessive, inherited, persistent, and unmanaged access are the four categories of agent access that create the most risk. How common is it for all four of those conditions to be present simultaneously, and what does each one look like when it becomes the vector for unintended behaviour?
Unfortunately, seeing all four conditions present at once is becoming increasingly common.
Excessive access occurs when an agent is granted more permission than its task requires. Rather than carefully scoping access, organisations often default to broad permissions for convenience, giving agents access to systems and data beyond their intended task.
Inherited access is often the most dangerous because it’s largely invisible. An agent operating through a human identity can inherit years of accumulated permissions by default, including access that may no longer be relevant to that person’s role.
Persistent access means permissions remain active long after the original task has been completed. The agent may no longer need access, but the credentials remain available, creating an ongoing risk if they are reused, misused or compromised.
Unmanaged access refers to credentials, tokens or service accounts that aren’t actively tracked, governed or owned by anyone. Organisations often don’t know they exist until something goes wrong.
Each of these creates risk individually, but together they create the perfect conditions for unintended behaviour. You end up with an agent holding broad permissions it doesn’t need, inherited access nobody realised it had, credentials that don’t expire, and limited oversight over how those privileges are being used.
4. An agent that uses legitimate capabilities in unintended ways is harder to detect than one that attempts clearly unauthorised actions. The actions are within the agent’s technical capability. What does detection look like in that scenario, and what monitoring architecture is required to catch it?
When an agent is operating within its authorised permissions, technically every action it takes is allowed. That means it doesn’t trigger the traditional security signals organisations are used to looking for. The issue only becomes visible when you look at the pattern, volume and sequence of activity in the context of the task the agent was meant to perform.
That’s where many organisations have a blind spot. They can see what an agent accessed, but they often can’t determine in real time whether that access was actually necessary or appropriate for the task at hand. As a result, they often identify issues hours or even days later. For an AI agent operating at machine speed, that’s a significant window for unintended actions.
To close that gap, organisations need to maintain continuous control over AI agents throughout the entire access lifecycle, rather than focusing solely on the point of authentication. That means having real-time insight into agent activity, understanding whether behaviour aligns with established norms for that specific agent, and being able to immediately constrain or revoke access when actions deviate from approved parameters. Together, these capabilities help organisations govern autonomous activity, reduce uncertainty around agent behaviour, and prevent security, compliance or operational risks from escalating.
5. From a C-suite perspective in APAC, how much of the accountability question is being addressed in AI deployment decisions right now, and what changes when an incident like the Medicare breach makes the accountability gap visible rather than theoretical?
From my perspective, accountability is still largely being treated as a procurement and policy question rather than an operational one.
Our research shows that nearly every organisation has policies governing how AI should be used, but policy alone doesn’t create accountability. Accountability comes from being able to prove who authorised an agent, what it accessed, what actions it took and whether controls worked as intended. Incidents like Medicare make that gap impossible to ignore.
6. Government and critical infrastructure are the environments where unintended agent behaviour carries the highest consequence. What is the gap between the governance frameworks those organisations have and what the risk profile of their agent deployments actually requires?
Government and critical infrastructure organisations generally have strong governance frameworks, but those frameworks were largely designed to govern how AI is developed, approved and deployed. The risk profile of AI agents is different because it centres on what those agents can access and do once they’re operating inside an environment.
That’s the gap. Governance frameworks are focused on managing deployment risk, while agentic AI increasingly requires organisations to manage access risk.
In high-consequence environments like government and critical infrastructure, that’s critical. It’s no longer enough to know an agent was approved. Organisations need to ensure agents have access only to the systems, data, and privileges they need, and be able to prove that access is continuously monitored and controlled.
7. The organisations that are deploying AI agents responsibly right now share specific characteristics. What are those, and what is the most common difference between them and those deploying fast without the right controls in place?
The organisations getting this right consistently do three things well.
First, they treat every AI agent as an identity with a clearly defined owner and accountability structure. Second, they grant access based on the specific task being performed and ensure that those permissions expire when the work is complete, preventing access from accumulating over time. Third, they maintain a clear audit trail linking sensitive actions back to the person who authorised the agent to act.
Just as importantly, they’ve planned for when something goes wrong. They have visibility into agent activity, the ability to intervene in real time, and a process of revoking access when behaviour falls outside expectations.
8. If you could identify one change in how enterprises deploy agents, how vendors build them, or how regulators govern them, what would it be and why isn’t it already happening?
If I could change one thing, I’d eliminate standing access for AI agents.
Agents should only receive the access required to complete a specific task, for the duration of that task, with those permissions automatically revoked once the work is complete.
Many of today’s AI risks stem from agents retaining access long after they need it or being granted permissions far broader than the task requires. Limiting access to what’s needed, when it’s needed, significantly reduces the potential impact when something goes wrong.
This will become even more important as Australia’s National AI Standards are introduced and anticipated AI legislation places greater emphasis on accountability and governance. Organisations won’t just need to show they have policies governing AI use. They’ll need to show those controls are enforced in practice, particularly as AI agents gain access to sensitive data and systems.
That means being able to prove who approved an agent, what permissions it was granted, what actions it took, and when its access was revoked. Ultimately, responsible AI deployment is about proving they only had access to do what they were authorised to do.
Bio:
Cynthia Lee is a cybersecurity professional at Delinea, specializing in identity security and privileged access management. She focuses on emerging security challenges, including the evolving role of artificial intelligence in cybersecurity. Through her industry insights and engagement, she highlights the importance of securing digital identities, managing access risks, and helping organizations navigate the evolving cybersecurity landscape.
Company:
Delinea is a leading identity security company helping organizations protect human, machine, and AI identities across modern enterprise environments. Its unified platform provides privileged access management, intelligent authorization, identity governance, and threat detection. Delinea enables businesses to reduce identity-related risks, strengthen cybersecurity, simplify compliance, and secure access across cloud and on-premises infrastructure.
