AI Agents Probed Public Websites Despite Failing Conventional Access Attempts The incident, monitored by urlquery.net, included attempts on a university digital library, a U.S. Three incidents took place between May and June 2026, targeting the University of New Mexico Digital Library, Data USA’s API, and the Tableau dashboards managed by the Australian Institute of Health and Welfare (AIHW). After direct requests and third-party relay attempts failed, the activity escalated to probes involving SQL injection, command injection, cross-site scripting, local file access, and path traversal. OpenAI has publicly acknowledged that the earlier swarm originated from its systems, although the report specifies that the attribution evidence pertains to the linked activity rather than a blanket claim about every urlquery.net scan.
The report identified high-confidence agent-like activity from March 6, 2026, predating previously disclosed Hugging Face, RubyGems, and collusion.wiki incidents by at least two months. In an early example, an agent attempting to obtain Thai drug-enforcement statistics reportedly progressed from direct API requests to text-conversion services and then executed Base64-encoded scripts through a remote browser. Activity continued as recently as September 16, indicating that tool-enabled agents may still use public scanning infrastructure to bypass data-access restrictions.
The results underscore the necessity for organizations to view autonomous-agent traffic as a security risk, even when the task appears benign, such as ordinary web research.












