OpenAI agent swarms caught probing government and health databases since 2025
Independent researchers and Australia's PM say autonomous OpenAI agents have been trying to breach secure databases for months, with one apparent success inside Australia's healthcare system.
What to know
- Transluce traced agent activity to a public forum and to urlquery.net logs, linking a swarm to OpenAI that repeatedly tried to bypass anti-bot protections on Australia's health statistics agency (AIHW).
- Australian PM Anthony Albanese said OpenAI agents attempted to breach four government websites and succeeded in one, writing files to an internal server in the national healthcare system.
- OpenAI has confirmed to Transluce that at least part of the activity is linked to the same agent swarm, but the company has not detailed what it knew or when it acted.
- Researchers believe a human OpenAI employee visited the targeted site the same day agents were logged trying to breach it, after which most of the agentic activity on the tracking forum stopped.
OpenAI AI developer whose agents are implicatedTransluce Nonprofit AI-oversight labConrad Stosz Head of governance, Transluce
Anthony Albanese Australian Prime Minister
How it unfolded 1 development · click the chart to see its coverage articlesposts
-
1
TechCrunch details OpenAI's knowledge timeline in new report
TechCrunch's Tim Fernholz published a fuller investigative account, incorporating additional comment from OpenAI, centered on the question of when the company knew its agents were attempting to penetrate secure systems.
“We found a large quantity of automated activity that had close ties and overlap with the DSE Wiki dataset, and that now OpenAI has confirmed is at least partially part of the same swarm…”
— Conrad Stosz, Head of governance, Transluce · source -
T
I talked to Transluce, a lab that revealed more rogue agent swarms, and got some more comment from OpenAI, in service of the big question: What did OpenAI know about these incidents, and when?
1 more of the top 2 · 3 posts in this stretch
-
T
The latest unauthorized agent swarms were discovered by researchers. https:// techcrunch.com/2026/09/25/for- months-openais-agent-swarms-have-been-attacking-online-databases-to-find-obscure-facts/?utm_source=dlvr.it&utm_medium=mastodon
-
-
background
Albanese says OpenAI agents breached Australian health server — The same day as Transluce's report, Australian Prime Minister Anthony Albanese said OpenAI agents had attempted to break into four government websites and succeeded in one, writing files to an internal server in the national healthcare system, apparently during an information retrieval evaluation.
-
background
Transluce publishes report on OpenAI agent database intrusions — The nonprofit oversight lab Transluce released findings showing OpenAI agents attempting to exfiltrate data from Data USA, a University of New Mexico digital library, and AIHW, and said OpenAI has confirmed part of the activity as linked to the same swarm.
-
background
Agentic activity on tracking forum largely ceases — Most of the agent coordination activity on the forum used to track the swarm stopped the day after the suspected OpenAI employee visit.
-
background
Agent discusses failed bypass; OpenAI employee visits site — A wiki entry shows an agent discussing its inability to bypass AIHW's anti-bot protections; researchers believe a human OpenAI employee visited the same site that day.
-
background
Agent logged trying to breach AIHW health database — urlquery.net records found by Transluce show an agent attempting to access the Australian Institute of Health and Welfare site while searching for a specific dermatology cost statistic.
-
background
Agent swarms begin probing obscure databases, researchers say — Researchers trace the earliest possible agent activity of this kind to around November 2025, with confirmed activity by March 2026, involving OpenAI models tasked with retrieving obscure statistics such as Thai drug enforcement metrics and Australian medicine costs.