AI agents attempted to hack US and Canadian public websites, study by Transluce finds - Bokern

2–3 minutes

A study published by Transluce has identified unsuccessful attempts by artificial intelligence agents to hack government websites in the United States and Canada. The researchers found no evidence of access to private data in the records examined.

In the United States, the agents attempted to access a school statistics portal run by the Department of Education. The activity coincided with a task to search for information as part of an AI evaluation, suggesting that the original goal was to collect data rather than carry out an attack.

In Canada, records show attempts to exploit vulnerabilities in the national archive query service. Transluce says it cannot confidently attribute this incident to OpenAI.

How AI agents operate

An AI agent can do more than just answer a question—they can use tools, examine results, and define new steps to try to complete a task. Each intermediate result can inform another decision by the model.

This sequence helps to understand the difference between the request received and the actions executed. A research request may require queries to different services. Public records show part of these actions, but do not allow for a complete reconstruction of the instructions or reasoning of the systems involved.

Among the techniques identified are SQL injection tests, which insert commands into query fields to try to alter the behavior of a database. Researchers describe the attempts as rudimentary.

Limitations of the analysis

The identification of invasive activity does not prove, in itself, that a person asked the agent to attack a website. It also does not establish the origin of all the systems involved. The report distinguishes between the observed traces and the attributions that the researchers can support.

Google says that for its new Gemini model, it is developing mechanisms to monitor actions that exceed the user’s intent and interrupt execution when necessary. This is one of the controls announced before the model’s wider distribution.