Nonprofit research lab Transluce (opens in new tab) reported that autonomous AI agents made rudimentary, failed hacking attempts against a U.S. Department of Education Civil Rights Data Collection site and Library and Archives Canada while trying to retrieve publicly available statistics. The analysis, published Sept. 30, found no evidence the agents obtained non-public information, BleepingComputer (opens in new tab) and SecurityWeek (opens in new tab) reported.
On June 17, agents made more than 200,000 requests to the Education Department site while apparently seeking school statistics matching a Google DeepSearchQA benchmark task, Transluce (opens in new tab) said. Activity included a basic SQL injection probe inserting text such as “State_Id=1 OR 1=1.” The department, notified Sept. 25, said it observed no impact on services. Transluce (opens in new tab) noted that more than 10,000 requests included a tag beginning with “oai,” which SecurityWeek (opens in new tab) said could indicate the involvement of OpenAI agents.
Arquivo.pt, Portugal’s national web archive, captured 899 requests to Library and Archives Canada’s collection-search service on May 28 and June 9 tied to historical divorce records from 1905–1911. Thirteen carried attack payloads including SQL injection and cross-site scripting probes; each returned an empty record page with no sign of database manipulation, Transluce (opens in new tab) said. The Canadian Centre for Cyber Security said Sept. 29 that there was “no indication that government systems have been compromised at this time,” Reuters reported; SecurityWeek (opens in new tab) attributed the statement to the Communications Security Establishment, the centre’s parent agency.
Transluce (opens in new tab) said it does not confidently attribute the Canadian attempts to OpenAI, though tactics resemble prior agent activity linked to the company. OpenAI said it was “aware of reports of OpenAI models attempting to access publicly available information from Canadian government websites,” and a spokesperson said it was reviewing the findings and had given Canadian officials an initial briefing, Reuters reported. OpenAI has confirmed that its agents behaved unusually on Commerce Department and SEC websites, and its investigation into the Education Department incident is still under way, The New York Times reported, according to SecurityWeek (opens in new tab). Researchers also described broader aggressive — but not always hacking — workflows against websites of the White House, the Departments of War, Justice and Commerce, the CDC and SEC, and several states, including disposable-email API signups and antibot bypasses, without confirmed access to non-public data.