Salt expands Agentic Detection and Response with LLM runtime and prompt security

Industry

Australia’s Legacy Stocktake Needs to Count What AI Agents Can Reach

November 16, 2026

Alexander Larcos
Regional Director, APAC

The Medicare statistics portal incident shows that an old system’s real risk sits in the interfaces it exposes. Agents are very good at finding them.

On 18 June, an AI agent working on a research task for OpenAI went looking for data on government health spending. It ended up with non-public access to the Medicare Statistics Reporting Service run by Services Australia. OpenAI has since said the agent “ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files.” It also said the agent used that access to review technical system information and source code for the service. The government says there is no evidence that personal patient records were accessed, and the portal has been taken offline.

Prime Minister Anthony Albanese summed it up: the agent “found a way around those blocks, didn’t accept ‘no’ for an answer.”

The agent's entry method has not been made public. The government released the 18 recommendations from its rapid review of the incident on 10 November, but it has not said publicly that the separate forensic investigation by Services Australia and the Australian Signals Directorate (ASD) has concluded. This post doesn’t speculate on the method. It looks at what the government’s response means for Australian organisations running legacy systems alongside AI agents.

Canberra’s answer is a legacy stocktake

The Department of Home Affairs has directed all non-corporate Commonwealth entities to complete a legacy technology stocktake and develop a legacy technology risk management plan by 31 March 2027. Agencies must establish targets for reducing legacy systems, document mitigations for systems they retain, and accelerate patching of critical vulnerabilities. However, agencies operating Systems of Government Significance face an earlier deadline: by 31 December 2026, they must implement prescribed risk-reduction measures and report completion of their legacy technology stocktake for those critical systems.

The need is well documented. In the ASD’s Commonwealth Cyber Security Posture in 2025 report, 59% of entities said legacy technology was affecting their ability to implement the Essential Eight. Government Services Minister Katy Gallagher said of the Medicare statistics portal: “It dates back decades.”

But a stocktake built the traditional way, as a list of servers, operating systems and end-of-support dates, will miss what an agent actually touches.

Agents don’t care how old a system is

An agent doesn’t interact with a system’s age. It interacts with whatever the system exposes: a portal, a query service, an API, an admin endpoint, a key left somewhere it shouldn’t be. As UNSW’s Professor Salil Kanhere told the Guardian, “A 15-year-old system that is properly supported, patched and properly isolated would perhaps present less risk than even a newer system that might not be properly maintained.”

The other incidents OpenAI disclosed make the same point. At the Victorian Agency for Health Information, OpenAI said its agent “discovered an exposed access key” that let it query a reporting system. At the Australian Institute of Health and Welfare, attempts to bypass access controls failed. The deciding factor in each case was not the age of the system. It was what the system exposed, and what could reach it.

Salt has written about this exposure for years. In a 2022 post, we grouped forgotten and unmanaged interfaces into three types:

  • Zombie APIs: endpoints that once served a purpose and were never switched off. They miss patches because nobody remembers they exist.
  • Shadow APIs: interfaces stood up outside formal governance, often to get a project out the door on time.
  • Ghost APIs: third-party interfaces that arrive inside packaged applications, appliances and SaaS services, and rarely make it into an inventory.

Legacy estates tend to collect all three. Over the years, older systems are extended with portals, reporting services and integrations so newer applications can use their data. Each one is a door, and an agent can find a door far faster than a person can.

Why agents change the maths

Attackers have gone after forgotten endpoints for as long as there have been endpoints. Agents add speed and persistence. They try, fail, adjust and try again at machine pace, and they can be pursuing a goal that looks entirely legitimate, like researching public health spending.

This isn’t the first time a capable agent has found a shortcut nobody intended. In July, OpenAI disclosed preliminary findings on the Hugging Face incident: models under internal evaluation had moved beyond their test environment and reached Hugging Face’s production database. Separately, Salt Labs showed that a single email could hijack the Manus AI agent platform (the issue has since been resolved). Its guardrail detected the attack, but only after the malicious code had run.

The pattern is consistent. The risk lies in what an agent can reach and do, and whether anyone can see it and stop it in time. Our State of Agentic AI and API Security 2H 2026 report, a survey of 383 security practitioners, found:

  • 49.8% have confirmed or suspected that an AI agent took an unintended or unauthorised action
  • Only 12.5% can trace the full path an agent takes across their systems
  • Only 17.5% can automatically stop an AI agent

It isn’t only a government problem

Banks, insurers, super funds, health providers, universities and critical infrastructure operators run the same mix of long-lived core systems and fast-moving AI projects. Several Australian obligations already reach the interfaces in between:

  • CPS 234: In our reading, APIs fall within what APRA’s information security standard treats as information assets, which must be classified, protected by tested controls, and evidenced. We covered this in detail in Under CPS 234, Your APIs Are Information Assets.
  • CPS 230: in force since 1 July 2025, it raises expectations on operational risk and on the third-party providers institutions depend on.
  • SOCI Act: responsible entities for certain critical infrastructure assets must maintain a risk management program covering the hazards to those assets.
  • Privacy Act: from 10 December 2026, organisations must explain in their privacy policies when computer programs use personal information to make decisions that significantly affect individuals. If agents are part of those decisions, you need to know where they are and what they touch.

Five questions to add to any legacy stocktake

  1. What does each system expose? List the APIs, portals, query services, and admin endpoints in front of every legacy system, including the undocumented ones. An inventory built from documentation alone will miss zombie and shadow endpoints.
  2. Which credentials can reach it? Map the keys, tokens, and service accounts, where they are stored, and who or what holds them. One exposed key can turn a protected system into an open one.
  3. Who and what is calling it? Separate human users, approved agents and unknown automation, and know what normal behaviour looks like for each.
  4. Can you follow the whole sequence? A single request can look harmless. Watch the chain of actions an agent takes across systems, and gate sensitive actions before they run, not after.
  5. Can you evidence it? Keep records that will stand up when the ASD, your auditors or APRA ask what happened.

Where a system can’t be replaced, the Australian Cyber Security Centre’s guidance is to segregate or isolate it. These five questions tell you what needs isolating and whether the isolation is holding.

How Salt helps

Salt Security is the leader in agentic and API security. The Salt Security Agentic Security Platform discovers the agents, MCP servers and APIs operating in an environment, including outdated and undocumented endpoints. It maps the connections and permissions between them and monitors agent behaviour at runtime, so security teams can see the path an agent takes and act when its behaviour departs from what was intended.

For Australian organisations, the platform runs in AWS’s Sydney region with data kept in Australia, backed by an in-country team. It won’t do your stocktake for you. It gives the agent and API part of it a live inventory backed by evidence, instead of a spreadsheet that is out of date before March.

Your agents are acting right now. See what they can reach.

Get a demo: salt.security/demo-request

Frequently asked questions

What happened with the OpenAI agent and the Medicare statistics portal?

On 18 June 2026, an OpenAI agent gained non-public access to Services Australia’s Medicare Statistics Reporting Service during a research task. OpenAI says it ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files. The government says there is no evidence that personal patient records were accessed. How the agent got in has not been disclosed.

What is the Commonwealth legacy technology stocktake?

It is a direction from the Department of Home Affairs requiring non-corporate Commonwealth entities to inventory their legacy technology and produce a risk management plan, including reduction targets and mitigations, by the end of March 2027. Entities operating Systems of Government Significance face an earlier deadline of 31 December 2026 for those systems.

What is a zombie API?

A zombie API is an endpoint that once served a purpose but was never decommissioned. Because it is no longer maintained, it often misses security updates and can give attackers, or AI agents, a way into the systems behind it.

Does CPS 234 apply to APIs?

CPS 234 does not name APIs, but APIs fall within what the standard treats as information assets. APRA-regulated institutions need to classify them, protect them with tested controls, and be able to evidence those controls.

Our latest posts