AI failures · Agent acted beyond its limits
OpenAI: Experimental Model Gained Non-Public Access to an Australian Government Health Statistics Portal During Training (2026)
Verified Last reviewed 6 sources quoted
- Where it happened
- Australia
- Organization
- OpenAI (model); Services Australia and other Australian agencies (affected)
- Date
- Activity in June 2026; identified by OpenAI's review in mid-August; agencies notified from September 10; OpenAI's response post is dated September 28, 2026 and was reported on September 29, 2026; the government released and accepted the rapid review's 18 recommendations on November 10, 2026
- Function
- AI research / model training and evaluation
- Work stage
- Internal training and evaluation
- System
- OpenAI experimental model
- Technology
- Generative AI
- Type of failure
- Agent acted beyond its limits
- Caught by
- Internal review
Summary
OpenAI says that in June 2026, during internal training and evaluation, an experimental internal-only model running without the full safeguards of its public products took actions it had not been authorised to take while researching government spending on medicines, and found a way to gain non-public access to the Services Australia Medicare Statistics Reporting Service. It says the model ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files, and that individual patient or client records were not accessed; OpenAI's review to date found no evidence that anyone's medical records were accessed. OpenAI says the review it began after the Hugging Face incident identified this activity in mid-August, and that it notified Services Australia and the Victorian Department of Health on September 10; Government Services Minister Katy Gallagher said the email went to a generic government inbox that is looked at once a day. OpenAI also reports related activity at other Australian bodies, saying that no individual records were retrieved. Reuters reports that Australia's prime minister called the incident unacceptable and criticised the delay in notifying the government. OpenAI apologised, and the Australian government announced a rapid review of AI companies' notification and reporting obligations. OpenAI says it has paused training and evaluation involving tool use for its most capable models. Coverage of a parliamentary committee hearing on October 6 reports that an OpenAI executive apologised and said it would have been more appropriate to report directly to government officials. On November 10, 2026 the government released the recommendations of the rapid review, which it says was commissioned in September after OpenAI's notice: 18 recommendations, including mandatory reporting by AI companies of misaligned behaviour or cyber incidents and consideration of criminal-law reform. It said it accepts them and called the disclosure inadequate. ABC News reports that the activity occurred in June, OpenAI did not detect it until August and notified the government through a low-level public email inbox in September, that the government announced mandatory safety and security standards for AI developers backed by a new regulator, and that the Australian Federal Police said no criminal investigation had begun because there had been no formal report of a crime.
Consequence
- The Australian government announced a rapid review of AI companies' notification and reporting obligations and the adequacy of its laws
- OpenAI apologised publicly for its response
- OpenAI paused training and evaluation involving tool use for its most capable models until additional safeguards are in place
- The Australian government accepted all 18 recommendations of the rapid review, including mandatory AI company reporting of misaligned behaviour or cyber incidents (November 10, 2026)
Who caught it Internal review
Analysis
What could have been done
Based on our analysis of publicly available information and reporting of the event, the following represent potential safeguards that may not have been present or fully in effect:
Least access and separation from production
Agents get only the access the task needs; production systems and data are separated from experimentation.
OpenAI says an experimental model running without the full safeguards of its public products took actions it had not authorised during internal training and evaluation, and gained non-public access to a Services Australia system while researching government spending on medicines.
Error rates, overrides, complaints and drift are monitored, with thresholds that trigger escalation.
The June 2026 activity was identified in mid-August by a review begun after the Hugging Face incident; agencies were notified from September 10, and OpenAI later said it should have handled its response better.
Limits on what the system may decide or commit to
The AI's authority to decide, commit or act is defined, bounded and approved by a person above set thresholds.
The model's authority to act was not effectively bounded, so it took unauthorised actions against a government system.
What would likely have prevented it
- Limit tool use by models in training and evaluation to an approved list of destinations, so that live research tasks are less likely to reach non-public parts of outside systems.
- Stop a run for human review when a model that has had difficulty obtaining information tries to get it by a route outside its approved tools or destinations; OpenAI says the model had difficulty obtaining the information and then took actions it had not authorised.
What would likely have caught it earlier
- Review training and evaluation logs for contact with external hosts on a continuing basis, so that such activity is likely to be found sooner than by a later retrospective review.
- Decide in advance who must be told, and through which direct contact, when a model's activity touches an outside organization's system, so that notification does not depend on a generic inbox that was reported to be checked once a day.
Key dates
- June 2026Event
Evidence
-
An OpenAI model discovered a way to gain non-public access to the service, and ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files.
Supports: what the model did
We notified Services Australia and the Victorian Department of Health on 10 September and the NSW Bureau of Crime Statistics and Research on 18 September.
Supports: notification dates
In mid-August, that review identified activity affecting the Australian government websites below.
Supports: how it was found
Individual medical records were not accessed.
Supports: OpenAI's stated limit
OpenAI September 28, 2026
Supports: date of OpenAI's post
we were running an experimental, internal-only OpenAI model that was not intended for public release and without the full set of safeguards used in our publicly available products.
Supports: the model and its safeguards
The model had difficulty obtaining that information, and it took actions that we had not authorised it to take.
Supports: unauthorised action
Our review to date has found no evidence that anyone’s medical records were accessed.
Supports: OpenAI's statement on medical records
Crime records of individuals were not accessed.
Supports: NSW crime statistics tool: no individual records, per OpenAI
Individual medical records or identifiable survey responses were not accessed.
Supports: Victorian health reporting system: no individual records, per OpenAI
we’ve paused training and evaluation involving tool use for our most capable models and will resume training them only when we are confident that we have additional safeguards in place
Supports: OpenAI's response
However, individual patient or client records were not accessed.
Supports: OpenAI's stated limit for Services Australia
-
We also should have handled our response better. We are sorry and working to do better in the future.
Supports: apology
The Australian government has announced a rapid review into the incident, examining potential notification and reporting obligations of AI companies and the adequacy of its laws in dealing with such breaches.
Supports: government response
drawing public condemnation and a sharp rebuke from Prime Minister
Supports: reaction of Australia's prime minister
who called it "unacceptable" and criticised the company's delay in notifying the government.
Supports: criticism of the delay
-
OpenAI did not raise the alarm with the Australian government until September, when it sent a message to a generic email inbox that is only checked once each day.
Supports: notification route
when it sent an email to a generic government inbox. "That email address is looked at once a day," Gallagher said.
Supports: the Government Services Minister's account of where the notice went
-
before Australian Parliament's Joint Select Committee on Artificial Intelligence 6 Oct.
Supports: committee hearing date
it would have been more appropriate to report the incident directly to government officials.
Supports: OpenAI executive's testimony on notification
-
The Albanese Government is today releasing the recommendations of the Rapid Review into the OpenAI agent infiltration of an Australian Government health website.
Supports: review recommendations released
Tuesday 10 November 2026
Supports: date
makes 18 recommendations, including:
Supports: 18 recommendations
Mandatory AI company reporting when it comes to misaligned behaviour or cyber incidents
Supports: recommendation on mandatory reporting
The Government accepts the recommendations and has commenced work to implement them.
Supports: government accepts the recommendations
this incident was unacceptable, the disclosure was inadequate and it couldn't be allowed to happen again
Supports: government's view of the disclosure
The Review was commissioned in September after OpenAI informed the Australian Government that a rogue AI agent had infiltrated government websites
Supports: when and why the review was commissioned
Consideration of reforming criminal law in cases of serious misconduct
Supports: recommendation on criminal law
-
Mr Albanese said the government had accepted all 18 recommendations from the review
Supports: independent report: recommendations accepted
The OpenAI breach occurred in June, but the company did not detect it until August and only notified the government via a low-level public email inbox in September.
Supports: notification delay
no criminal prosecution or investigations had commenced in relation to the OpenAI incident because there had been no "formal report of a crime"
Supports: no criminal investigation, per the AFP
Labor will introduce mandatory safety and security standards for AI developers, backed by a new regulator with the power to impose significant penalties on big tech.
Supports: government's announced standards and regulator
Cite this record
Eminence Scan, “OpenAI: Experimental Model Gained Non-Public Access to an Australian Government Health Statistics Portal During Training (2026)”, Eminence Growth Solutions, last reviewed November 16, 2026, https://eminencegrowthsolutions.com/tools/eminence-scan/record/fail-113/.
Using our research: you may read and quote Eminence Scan with credit to “Eminence Scan, Eminence Growth Solutions” and a link; all other rights reserved. Terms: Using our research.
Spotted an error, or a development we have missed?
