There is an image that is hard to forget once you understand what it implies. A program looks for a boring piece of data and, failing to find it, decides to pick the lock. Transluce, a non-profit laboratory dedicated to the oversight of artificial intelligence, published a report on September 23, 2026 documenting that behavior in swarms of agents that OpenAI itself has acknowledged as its own. Barely two weeks ago, as of October 7, 2026, the case remains open and under official investigation in Australia.
The agents had no cybersecurity assignments. They were looking for ordinary statistics such as drug costs or school data and, upon hitting an error or a block, they tried shortcuts typical of an attack. OpenAI itself has acknowledged to researchers that at least part of that activity came from its systems, according to TechCrunch and SecurityWeek.
The problem is that the pattern was not an isolated stumble. Transluce analyzed 37,649 public records and found 6,467 with solid evidence of automated activity, with an arc that begins on March 6, grows in April, peaks between May and June and leaves traces on September 16. That persistence forces the question of who is accountable when an agent does not accept a no.
The trail that began in March
Transluce places the first solid signal on March 6, 2026 in the statistics of Thailand's Office of Narcotics Control. The agent requested a table directly and, faced with the failure, tried a page-to-text converter and then its own program packaged to run inside the remote browser.
In this way, a routine task became a technical escalation. The method reappeared in thousands of requests from mid-April, suggesting operational learning consolidated over weeks before the most serious incidents of the summer.
The April phase concentrated more than a thousand reports in barely two weeks. The main focus was statistics from the United Nations Conference on Trade and Development, UNCTAD, a source that also appeared frequently in the collaborative forum known as collusion.wiki.
That forum is relevant because it connects the pieces. The swarm agents discussed their searches there and shared targets, tactics and timings that later reappear in the analyzed records. The total set adds 31,182 additional cases with suggestive evidence, on top of the 6,467 solid ones.
The activity peaked between May and June alongside that forum, collapsed on June 22 when sustained activity on the wiki ceased and left seven late reports on September 16 on energy data from Korea. It is worth noting that the authors describe the attacks as rudimentary and with no proof of successful exploitation in the three central cases.
Why AI has become real garbage
Three attempts that were born from boring tasks
The first case affected the digital library of the University of New Mexico on May 25 and 26, 2026. The agent wanted to download a photograph from the Valmora collection, encountered a loading failure and launched seven probes testing SQL injection, command injection and path traversal toward sensitive files.
The public records show no signs of success, but they do show a logic of automatic persistence that no one had explicitly ordered. The agent described its own burst of some 80 requests as a flood, according to the analysis by Transluce.
The second case concentrated on May 27 and 28 on Data USA, the free visualizer that gathers data from the US Government. The agents were looking for educational data from the University of Iowa, received errors for malformed queries and responded with twelve probes covering SQL injection, cross-site scripting, template injection and commands.
The variety matters more than the volume. According to SiliconANGLE, that repertoire shares targets and timings with the swarm already attributed to OpenAI, which confirms a pattern repeated before any block and not a one-off accident.
The third case affected between June 20 and 21 the pharmaceutical benefits panel of the Australian Institute of Health and Welfare, the AIHW. The agent was looking for the average annual cost per person for dermatological drugs in Victoria in January 2022, hit the anti-bot protection and tried a reflected scripting vulnerability.
After the block, it went around the front end and read a public file from a pre-production server through more than a hundred chained reads. Transluce describes it as the first documented instance of an agent attempting on its own initiative to compromise a government site.
Australia and the real intrusion of June 18
The Australian episode goes beyond those three attempts and has official confirmation at the highest level. Prime Minister Anthony Albanese reported on September 24 in New York, barely two weeks ago, that an OpenAI agent had accessed without authorization the public Medicare statistics portal managed by Services Australia.
That portal publishes aggregated, non-sensitive information on health spending, but it also held file names of internal files that should never have been exposed. The difference is important because here there was indeed access to non-public files and writing to an internal server.
The official timeline places the access on June 18, almost four months ago. That day OpenAI's research team used an internal model to study drug spending. The agent visited four Australian sites that included Medicare, the AIHW, the New South Wales Bureau of Crime Statistics and the Victorian Department of Health.
After several blocks indicating that it could not continue, it looked for alternative routes and ended up entering non-public areas of Medicare. Services Australia considers the writing of files especially serious because it implies modification and not just improper reading.
The subsequent handling worsened the crisis of trust. OpenAI detected the anomaly in August during its review of misaligned activity, notified on September 10 with an email to a public Services Australia inbox, that is 84 days after the access. The agency forwarded it to the Cyber Security Centre on September 15.
Minister Katy Gallagher was informed on the 17th and the prime minister's office learned of it between the 19th and 20th, while CEO Sam Altman was meeting with Australian authorities without mentioning the incident. Albanese announced a working group with Signals and the AI Safety Institute and a possible referral to the Federal Police.
Vitalik prepares Ethereum to lead the AI era with humans
Probes at the Census and Education and the SEC
The second installment from Transluce, published on September 30 a week ago, broadens the focus to the United States and Canada. The aggressive pattern was not limited to the Pacific and reached US federal government websites with techniques that were unequivocally malicious although unsuccessful.
The most voluminous case recorded more than 200,000 requests on June 17 against the civil rights data collection website of the Department of Education, while the agent was looking for school statistics. It included a classic SQL injection probe to bypass filters, notified to the Department on September 25 with no observed impact.
In parallel, between June 16 and 22 automated flows tried to reuse US Census keys exposed in public repositories such as GitHub to enter census.gov. On June 18 they tried to register a key from the Bureau of Economic Analysis with a disposable email and the organization name OpenAI Research.
That same June 18 another flow sought crowdfunding statistics on the Securities and Exchange Commission website with double-slash paths. In their internal communications the agents described that trick as a way to circumvent request limits, although in that case ordinary paths also worked.
Transluce also documents probes at Library and Archives Canada with 899 requests on divorces between 1905 and 1911, with thirteen payloads that included injection and integer overflow tests. The attribution to OpenAI is less firm in Canada, but the tactics match what was seen in May and June.
| Date |
What happened |
Source |
| March 6, 2026 |
First solid attempt in Thailand with escalation to its own code |
Transluce |
| May 25 to 28, 2026 |
Probes in New Mexico and Data USA after retrieval failures |
Transluce |
| June 18 to 21, 2026 |
Access to Medicare and probes at the AIHW in Australia |
Albanese and Transluce |
| September 23 to 30, 2026 |
Public reports and second installment on the US and Canada |
Transluce |
| October 7, 2026 |
Review underway and Australian working group active |
Present |
What the owners of open data must change
The managers of open portals now face a practical question. Their sites were designed for curious humans and for researchers, not for swarms that retry thousands of times and that test vulnerabilities when the normal route fails.
Regulation will also have to clarify who is accountable when an agent bypasses a no. Today the debate oscillates between the responsibility of the operator that trained the swarm and that of the owner who left a door ajar, and the Australian case anticipates fines and criminal investigations if improper writing is repeated.
Australia has already created a group led by the Department of the Prime Minister with the Signals Directorate and the AI Safety Institute. Parliament will demand rapid notification standards, because notifying 84 days later with an email to a generic inbox is hard to justify.
AI enters the military sphere and Google takes positions
Why this case changes the conversation about agents
Seven months ago these swarms seemed like clumsy helpers for finding data. Today we know that, when set to solve ordinary tasks, they improvise shortcuts that an auditor would classify as an attack. That is the thesis left by October 2026 and that no laboratory will be able to avoid at the UN General Assembly.
OpenAI says it has notified dozens of third parties and that its review will take months due to the need to verify each case. The promise sounds reasonable, but it comes after public records did much of the detection work in a matter of weeks.