More than 200,000 requests in one day on a U.S. Department of Education website, injection probes on Canadian archives: Transluce documents new excesses of agents.
The Transluce laboratory is continuing its investigation into "out-of-control" AI agents. Based on logs from the Portuguese web archive Arquivo.pt and the urlquery.net service, it describes two rudimentary and failed intrusion attempts. On June 17, agents who appeared to be looking for school statistics sent more than 200,000 queries to the U.S. Department of Education's civil rights data collection site, including a "1 OR 1=1" SQL injection probe. The data in question corresponds to a question from Google's DeepSearchQA benchmark, suggesting that the agents were evaluated on a document search task, not a hacking task. More than 10,000 queries had a label starting with "oai."
On May 28 and June 9, 899 requests were made to Library and Archives Canada's search engine for divorce records from 1905 to 1911; 13 of them contained attack charges (three SQL probes, five attempts to manipulate the output format). Transluce likens these tactics to activities that he had already attributed to OpenAI, without stating it here with certainty.
No non-public data was reportedly obtained. The department, notified on September 25, says it has seen no impact; the Canadian Centre for Cyber Security reacted publicly on September 29. The report also lists more aggressive, non-hacked polls on the websites of the White House, several departments, the SEC, the CDC and five states.
Read the full article on the original source
Go to the source →