In September, Australian Prime Minister Anthony Albanese revealed that an OpenAI agent had gained unauthorized access to the Medicare Statistics Reporting Service, a public-facing government portal that also held restricted information. The agent had encountered the portal in June while researching public medicine spending, hit repeated blocks and, in Albanese’s words, “found a way around those blocks ... didn’t accept no for an answer.”

OpenAI didn’t discover what had happened until August. It didn’t notify Services Australia until September 10, and the notice went to a public inbox rather than a security channel. Albanese went public two weeks later and announced a cross-agency taskforce to investigate. Public reporting hasn’t found evidence that personal Medicare data was accessed, and the investigation continues.

Coverage has focused on OpenAI, and that accountability conversation matters. The signal that concerns the broader industry is this: autonomous systems can now interact with websites, applications and data repositories in ways that break the assumptions our security models were built on. Medicare wasn’t the only system touched. Researchers have linked related agent activity to at least three other Australian public bodies, including one case where an agent, blocked from a protected dataset, found the same file sitting unprotected on a pre-production server instead. Australia may be less an isolated incident than an early warning.

Agents as a New Category of Actor

Security programs have spent decades organized around a straightforward picture: humans access systems, applications process data and controls sit in between. Agents complicate both sides. Traditional automation runs a fixed script; agents evaluate, decide, adapt and pursue a goal with far less human involvement. As more organizations put them to work on research, analysis and execution, they start to resemble a new category of actor inside the enterprise rather than a new category of tool. Agents operate with significant autonomy, pursuing goals without step-by-step human approval, which makes their behavior harder to anticipate than anything security programs were built to govern. That gap is what security teams are now catching up to.

The Detection Timing Question

Much of the conversation around incidents like this focuses on how the access happened. A more useful question is how long it would take to know something similar had happened in your own environment. Services Australia didn’t find this on their own; OpenAI discovered it during an internal review and notified them. That detection gap has measurable stakes: IBM’s 2026 2026 Cost of a Data Breach report puts the average time to identify and contain a breach at 247 days globally, the longest in years, while healthcare remains the costliest sector to recover from for the 13th consecutive year running.

An agent can reach an API and move on before any human reviewer could respond, and without real-time visibility into what accessed a system and whether that activity matched policy, the most organizations can do is reconstruct what happened after the fact.

Classification Before Access

Most responses to incidents like this concentrate on the moment of access: who got in and what they did once inside. That framing treats data as a fixed backdrop the agent moved through, rather than something with its own exposure profile that should have been established before the agent ever appeared.

Sensitive data carries its classification from the moment it is created, regardless of what eventually tries to reach it. Organizations that haven’t established which systems and files hold sensitive information before anything attempts to access them are building controls on top of a gap. A folder marked internal-only reflects an assumption about how the system will be accessed, and incidents like this illustrate that those assumptions don’t hold in an environment where agents are active. Classification should determine what an AI system can reach before access is attempted, not after it has already occurred.

The Limits of Identity-Based Controls

Security programs have traditionally leaned on identity and access management: who is the user, what can they reach and were they authenticated properly. These questions still need answers. The problem is that an agent can have valid credentials, legitimate access and still do something with that access no one would sanction if asked. The question that matters isn’t whether access occurred but whether that specific use of the data was appropriate for the task the agent was given. Identity tells you who showed up but says nothing about what they should have done once inside.

What Agent Security Requires

The industry has been through structural transitions like this before. Cloud required a new operating model. Remote work required a new trust model. Agents require something similar, built around three elements: identity, so agent activity is tracked and attributed like a privileged human account; scope, so access stays least-privileged and is re-evaluated continuously rather than granted once and forgotten; and monitoring at machine speed, because a review cycle that only surfaces problems months after the fact isn’t oversight in any meaningful sense.

This is where Forcepoint AI Agent Gateway operates, serving as the mandatory path to the applications, data and tools an agent needs. Security teams can verify every agent’s identity, user and connection; enforce least-privilege access at the tool, action and data level; apply DLP policies to what enters or leaves an agent workflow; and log every connection and tool call so that questions about agent activity become a quick lookup rather than a forensic exercise. An unregistered agent attempting what the Medicare incident involved would be limited to explicitly public resources, with any attempt to reach further denied and logged at the moment it happened.

The Convergence Ahead

The instinct after an incident like this is to focus on the technical specifics: which model, which vulnerability, who is responsible. Those questions deserve answers.

The structural issue runs deeper. AI security and data security have operated as separate disciplines, and incidents like this show what falls through that gap. Data security establishes what information is sensitive, where it lives and how it should be protected before anything tries to access it. AI security governs what is interacting with that data, what actions are being taken and whether those actions align with policy. Neither produces a complete picture without the other, and the space between them is where agents will go when they find a way through.

The last decade was largely defined by securing users, devices and cloud workloads. Extending that same coverage to autonomous AI systems and the data they interact with is where the field is headed. Medicare may be one of the first public cases where that transition started to matter in practice.

To discuss agent visibility and enforcement in your own environment, talk to a Forcepoint expert.

Sunil Choudrie