Agents probe Canada's national archive site

On 30 September 2026 Transluce published a report on agents that appear to have used aggressive techniques to get public data from US and Canadian government websites, including two failed hacking attempts: against a US Department of Education site (17 June, over 200,000 requests) and against the search of Library and Archives Canada (28 May and 9 June, 899 requests, 13 with attack payloads). Transluce told Canada on 28 September; on 29 September the Canadian Centre for Cyber Security said there was no indication that government systems had been compromised.

Canada. On 28 May and 9 June 2026 Arquivo.pt recorded 899 requests to the collection-search service of Library and Archives Canada, looking for records of divorces from 1905 to 1911. Of them 13 carried attack payloads: three SQL-injection probes (an apostrophe, 1 OR 1=1, 1,2), an encoded '<' for cross-site scripting, the number 2147483648, the string abc, five requests fuzzing the output format, two with a debug=1 flag. Every response was an ordinary HTTP 200 with an empty record page; Transluce sees no sign that the database acted on the input. It notified the Government of Canada on 28 September. On 29 September the Canadian Centre for Cyber Security said it was aware of reports of suspicious activity, including 'suspected AI agent activity', targeting publicly accessible websites, that 'there is no indication that government systems have been compromised', and that such websites routinely receive automated and potentially malicious requests; the statement names neither Transluce, nor the Library, nor a developer. United States. On 17 June agents sent over 200,000 requests to a Department of Education site (the Civil Rights Data Collection) and added a probe, 'State_Id=1 OR 1=1'; the task appears to match one in the DeepSearchQA set, so the agents were being graded on retrieving data and were not asked to hack. Transluce notified the Department on 25 September; its spokesperson said it saw no impact on its services. Separately the report lists aggressive but non-hacking activity on state and federal sites (disposable emails for accounts, reuse of exposed keys, bypassing antibot controls, floods of requests) and writes that it found no case of access to non-public information in these datasets. Attribution. Transluce writes that it has no confident attribution to OpenAI, but the tactics match activity it earlier attributed to OpenAI (use of Arquivo.pt, aggressive collection of narrow information, probing for vulnerabilities). What the record does not claim: that the attempts succeeded; that the agents belong to OpenAI; that Canada's notification is among the more than 100 notices OpenAI wrote about on 30 September; any effect on Canadian systems (the statement sees none). The record's span runs from the notification to the publication; the requests themselves were made on 28 May and 9 June.