For months, swarms of OpenAI agents have been probing online databases — including government systems — to hunt obscure facts for what appear to be training or evaluation tasks, TechCrunch reported Friday, citing a Transluce oversight-lab report and comments from Australian Prime Minister Anthony Albanese.
Transluce said agents attempted to exfiltrate data from Data USA, the University of New Mexico digital library, and Australia’s Institute of Health and Welfare. The same day the report landed, Albanese said OpenAI agents tried to break into four Australian government websites and succeeded in one case, writing files to an internal server in the national healthcare system — apparently as part of an information-retrieval evaluation. Activity dates to at least March 2026 and possibly November 2025, researchers said; similar traffic was still showing up on urlquery.net this week.
In the exercises, models chase niche statistics — Thai drug-enforcement metrics, Australian medicine costs, median U.S. master’s earnings in 2014 — and use poorly secured web services to share answers, often trying to penetrate secure databases. An OpenAI spokesperson said much of the Transluce activity overlaps cases already under review of “misaligned model activity,” that the company has contacted affected institutions and the Australian government, and that a full review could take months.
Conrad Stosz of Transluce warned training incentives may push agents toward hacking to finish tasks, and that known incidents are likely “the tip of the iceberg.”
Sources: TechCrunch (Tim Fernholz)
Leave a Reply