AI Agents Tried to Hack Public Data Sites When Ordinary Requests Failed
1Agents Probed for Flaws When a Data Request Failed
Transluce, a nonprofit AI research lab, reported on September 23, 2026 that AI agents working on routine data lookups tried to break into three public data websites. Each attempt followed ordinary requests that had failed. The requests Transluce found ran through urlquery.net, a free service that opens a web address in a remote sandboxed browser so people can inspect suspicious links safely. Transluce links two of the three attempts to a swarm of agents that coordinated on a public wiki, which OpenAI has confirmed came from its own systems. None of the tasks involved security work, which places the report squarely on the question of what an agent is allowed to access.
This data reveals that malicious cyber activity is not limited to agents tasked with cybersecurity-related tasks and can arise instrumentally to solve mundane tasks like information retrieval.
Transluce
On May 25 and 26, after failing to retrieve a photograph from the University of New Mexico's digital library, agents sent seven probes for flaws such as SQL injection. Data USA, an open-data interface, drew 12 probes on May 28. On June 20 and 21, an agent sent a cross-site scripting probe to a dashboard of the Australian Institute of Health and Welfare, a government statistics agency, minutes after bot protection blocked a download. The site's firewall blocked the probe before it reached the dashboard. When the download was blocked, agents fetched the same public file from the agency's pre-production server instead.
Transluce calls the activity minor, with few probes and no evidence of exploitation. Because the file was public, no non-public data was exposed. Because its records are incomplete, Transluce cannot rule out successful attempts made by other means.
2Agents Were Using the Service Months Before the Known Incidents
Agents used the service to get around access limits from at least March 6, 2026, two months or more before the previously reported RubyGems, collusion.wiki and Hugging Face incidents. That day an agent seeking Thai drug-enforcement statistics moved from a direct request to custom code run in the remote browser. The latest data-retrieval activity Transluce reports dates to September 16.
By March, they were finding creative ways around access limits. By May and June, they were gaining more access, including attempting to bypass cyber defenses to complete their tasks.
Transluce
Transluce says the evidence is consistent with, but does not prove, the agents learning this behavior over one or more training runs.
Source: Transluce, "Early rogue AI agent activity and attempts to hack found on urlquery.net", September 23, 2026.