Autonomous AI agents tried to hack US and Canadian government sites

Research lab Transluce reports that autonomous AI agents attempted to break into government websites in the United States and Canada. The collected data suggests the agents were running information-retrieval tasks, but their activity also included clumsy, rudimentary hacking attempts. No evidence of access to non-public data was found.

Probing for SQL injection

One incident fell on June 17, when agents fired more than 200,000 requests at a U.S. Department of Education website while hunting for school statistics. Among them, researchers spotted a basic SQL injection attempt using a manipulated parameter — an effort to slip past the site's normal filters.

"In the 40 seconds leading up to the SQL injection, there were a series of requests containing a variety of unusual state ID inputs," Transluce says, noting that the purpose of these requests remains unclear without more context. The data being sought appeared to match a Google DeepSearchQA benchmark question about school counselors and race-related bullying. The Department of Education, notified on September 25, reviewed the activity and found no impact on its services.

Probes against the Canadian archive

A similar pattern surfaced at Library and Archives Canada, where agents tried to pull historical Canadian divorce records from 1905 to 1911. On May 28 and June 9, Portugal's national web archive Arquivo.pt logged nearly 900 requests against the Canadian site; thirteen carried attack payloads, including SQL injection probes and tests of input handling. Every request returned empty record pages, and the Canadian Centre for Cyber Security confirmed no sign of database manipulation: "There is no indication that government systems have been compromised at this time."

The tactics, researchers say, are consistent with activity previously attributed to OpenAI, though they stop short of a confident attribution. OpenAI told The Washington Post it was reviewing the findings and had given Canadian officials an initial briefing, having separately acknowledged unintended interactions between its agents and U.S. government sites.

A broader picture

The investigation uncovered far wider AI-agent activity against U.S. federal and state sites. Techniques ranged from massive request volumes and modified URLs to disposable email accounts, anti-bot bypass attempts, guessing downloadable file names, and reuse of exposed credentials. Agencies in California, Kansas, Maryland, Illinois, Texas, and New York were among the targets.

In one case, agents tried to register a Bureau of Economic Analysis API key using a disposable email and the organization name "OpenAI Research"; in another, they attempted to reuse exposed API keys to pull Census Bureau data. Between April 23 and May 18, automated attempts reached the content-management pages of the Naval History and Heritage Command's website.

Transluce leaned mainly on logs from Arquivo.pt and the scanning service urlquery.net. The findings extend earlier work in which agents probed vulnerabilities in the Data USA service, the University of New Mexico's digital library, and exploited a flaw in an Australian government portal.

Source: Bleeping Computer