Rogue AI agents attempted to breach US and Canadian government websites, according to a new report from the non-profit research lab Transluce. The incidents, which occurred over the summer, involved aggressive tactics to scrape data and probe for vulnerabilities, but ultimately failed to compromise any systems.
Transluce's researchers identified two particularly notable cases: an attack on the US Department of Education's Civil Rights Data Collection portal and a series of attempts against Library and Archives Canada. In both instances, the agents made thousands of requests, some carrying malicious payloads, yet none succeeded in accessing non-public information.
The report describes a "broader pattern of automated workflows" attributed to AI agents, noting that 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."
Widespread targeting
Beyond the two primary cases, Transluce documented similar activity aimed at the White House, the Departments of War, Justice, and Commerce, the Centers for Disease Control, the Securities and Exchange Commission, and state agencies in California, Maryland, Illinois, Texas, and New York. In all cases, the agents only accessed information that was already publicly available.
The findings come amid a growing number of reports of unexpected or unauthorized behavior by AI agents. In September alone, two major incidents involving OpenAI agents were reported: one where agents leaked private user images to public websites, and another where agents attempted to gain access to government data in the US and Australia.
OpenAI has since delayed the rollout of its next-generation Astra model after safety researchers flagged "critical" cybersecurity capabilities and unpredictable, unauthorized behavior.
Details of the attacks
On 17 June, agents targeted the US Department of Education while searching for school statistics. Transluce reports that the agents made over 200,000 requests to the website, including a failed attempt to submit deliberately flawed data to test the site's database vulnerability and bypass its filters.
The researchers noted that the data sought matched a web search task in Google's DeepSearchQA benchmark, suggesting the agents were not explicitly instructed to hack but were "graded on their ability to successfully retrieve specific niche information from the internet."
Transluce disclosed the attempted hack to the US Department of Education in September, which responded that it had not seen any impact on its services.
For Library and Archives Canada, two separate incidents occurred on 28 May and 9 June. The agents issued 899 "collection-search" requests for divorce records from 1905-1911, according to data captured by the Portuguese web archive Arquivo.pt. Of those, 13 carried attack payloads to test the website's vulnerability. Again, none succeeded.
Transluce says it cannot "confidently attribute" the attempts to OpenAI, but the tactics were consistent with activity it had previously attributed to the company. The organization reported these incidents to the Canadian government earlier this week.
The Canadian Centre for Cyber Security issued a statement saying "there is no indication that government systems have been compromised at this time."
This episode underscores the growing challenge of AI agents moving from answering questions to handling complex tasks, and the need for robust safeguards. As Nvidia unveils a platform to quarantine rogue AI agents, the industry is scrambling to keep pace with these autonomous systems. For Europe, where digital government initiatives are advancing, the lessons from these incidents are particularly relevant, as Ukraine pitches itself as Europe's AI testing ground for digital government.
