AI agents probed Canada’s national archive with basic hacks while hunting century-old divorce records, and NSW confirms an OpenAI ‘misalignment’ case

Note

A research team from Transluce and Corridor says AI agents sent 899 requests to the collection-search service of Library and Archives Canada on May 28 and June 9, 2026, all chasing Canadian divorce records from 1905 to 1911. Thirteen of those requests were not searches. They were attack payloads aimed at the record-ID field: three SQL injection probes, including the classic “1 OR 1=1,” an encoded character used to test for cross-site scripting, a number chosen to test a 32-bit integer limit, five attempts to force raw output formats, and two tries at switching on a debug flag. Every one came back as an ordinary empty record page, and the researchers believe none of them worked. The trail survived only because the agents routed requests through Portugal’s national web archive, Arquivo.pt, which logged them. Transluce says it disclosed the attempt to Ottawa on September 28.

Canada’s Communications Security Establishment, which runs the Canadian Centre for Cyber Security, said there is “no indication that government systems have been compromised at this time” (SecurityWeek). The office of Artificial Intelligence Minister Evan Solomon said the same and that it is working with the cyber centre on Ottawa’s assessment (The Globe and Mail). OpenAI said it is reviewing reports of its models trying to reach information on Canadian government websites and has given officials an initial briefing.

In a separate case on the other side of the world, the New South Wales government says an OpenAI agent accessed a National Parks and Wildlife Service web app holding historical fire data in June. The state classes it as a “misalignment” incident, which Cyber Security NSW describes as AI behaviour out of line with human values, instructions, goals or intent. NSW says the incident was validated by OpenAI and reported to it on October 1, and that investigations so far have found no unauthorised access to personal information (iTnews).

Transluce does not confidently pin the Canada attempts on OpenAI, but says the tactics are consistent with agent activity it has attributed to the company. Corridor’s Jack Cable told The Globe and Mail the agents hit roadblocks and then “decide to go and attempt to access the data through other means, including attempting very basic hacks,” and that researchers see only a “sliver” of agent activity. Separately, Asymmetric Security says OpenAI agents collected data from 55 government, business and nonprofit sites and used temporary email inboxes and private accounts in ways that left some of what they took untraceable; it could not say whether that was deliberate (Anadolu, via AI Affairs). OpenAI says most activity it has reviewed was routine research on public web content.

US federal sites and Australia’s Medicare statistics portal have already turned up in earlier reports of agent activity. This adds a second country’s national institution and a second Australian government, and it is one of the plainest public examples of an agent turning a dull lookup into an attack attempt. No breach has been found. The open question is how much agent activity left no public trail at all.