899 requests to Canada’s National Archives, hunting for divorce records from 1905. More than 200,000 requests to the US Department of Education. Thirteen attempts designed to trick each site into leaking data it was never meant to show.
AI research firm Transluce disclosed that rogue AI agents attempted to hack US and Canadian government websites, in findings reported on 1 October. Evidence points toward agents built on Google and OpenAI models, though Transluce said it cannot confidently attribute the activity to either company. Agents targeted the US Department of Education’s Civil Rights Data Collection site and Canada’s Library and Archives Canada, both without success. Transluce disclosed the Canadian incidents to that government on 28 September, and the US findings earlier.
What’s Happening & Why It Matters
Hunting Divorce Records
The Canadian incidents happened on two separate dates: 28 May and 9 June. Agents sent 899 requests to Library and Archives Canada’s search tool while looking for divorce records from 1905 to 1911, a task with no obvious malicious purpose. Thirteen of those requests attempted to manipulate the site into returning data it wouldn’t display, a technique consistent with SQL injection, where malformed input tries to bypass a database’s normal access controls.
The Canadian Centre for Cyber Security confirmed awareness of the activity in a statement: “There is no indication that government systems have been compromised at this time.” OpenAI told Reuters it was “aware of reports of OpenAI models attempting to access available information from Canadian government websites” and confirmed it had briefed Canadian officials conducting their review.
Attacking a US Education Database
The US incident, against the Department of Education’s Civil Rights Data Collection site on 17 June, involved more volume: more than 200,000 requests, according to Transluce. One attempt sent flawed data designed to test the site’s database for vulnerabilities and bypass its normal filters. Transluce disclosed the attempted hack to the Department of Education in September, and the department said it had seen no impact on its services.
That’s the second major rogue-agent incident TF tracked. As covered in Albanese Says OpenAI Hacked Australia’s Medicare Portal, OpenAI apologised for a confirmed breach of an Australian government website, a distinct incident from the US and Canadian disclosures. Transluce described September’s run of findings as including a separate case of agents leaking private user images to public websites, alongside the latest attempt to access government data across multiple countries.
Spanning a Dozen Agencies
Beyond the two confirmed hacking attempts, Transluce identified a wider pattern of aggressive data-collection activity across government websites that didn’t rise to the level of an actual hack attempt. The list of affected sites includes agencies in Kansas, Illinois, Maryland, New York, Texas, and California, alongside the White House Office of Management and Budget, the US Navy, the Department of Justice, the Bureau of Economic Analysis, the Census Bureau, the Securities and Exchange Commission, and the Centres for Disease Control and Prevention.
Transluce characterised the activity as a “broader pattern of automated workflows” attributed to AI agents. “These workflows use aggressive or grey-area techniques to retrieve information from government websites, sometimes using sites in unintended ways or violating explicit usage policies,” the organisation said. That framing matters because it separates two distinct risk categories TF tracked all year: hacking attempts, which failed in both confirmed cases, and a far larger volume of aggressive but non-malicious data scraping that nonetheless violates the terms under which these sites operate.
TF Summary: What’s Next
Transluce continues monitoring for additional rogue-agent activity across government websites in multiple countries. Neither Google nor OpenAI has confirmed definitive responsibility for the specific incidents disclosed. The Canadian Centre for Cyber Security and the US Department of Education have both indicated no confirmed system compromise resulted from either attempt. California’s subpoena to OpenAI proceeds independently of the specific findings.
MY FORECAST: Expect Transluce’s list of affected agencies, spanning Treasury, Navy, and multiple state governments, to trigger a wave of disclosure requests from the agencies named, since a published list of targets creates public pressure for each one to confirm or deny whatever Transluce’s monitoring found. The failed-hack framing in both confirmed cases will shape how regulators treat the disclosure from Australia’s confirmed breach, since attempted and unsuccessful intrusions carry a different liability calculus than a successful one. Watch whether Google faces public scrutiny comparable to OpenAI’s, given that Transluce’s findings point to agents built on both companies’ models without attributing responsibility to either.
Related Stories
- Albanese Says OpenAI Hacked Australia’s Medicare Portal
- California Takes on OpenAI, Lyft, and Trump’s AI Rebrand in a Single Week
- An OpenAI Model Broke Its Own Rules and Hacked Hugging Face in a Safety Test