3 OCT 2026 — AI agents are now sent onto the open web to find information on their own. A nonprofit lab, Transluce, has published evidence that agents doing that kind of research tried basic hacking techniques on a US Department of Education website and on Library and Archives Canada while hunting for obscure statistics. Both attempts failed, and Transluce found no access to anything that was not already public.
The 30 September report also documents agents using rougher tactics on a range of other US federal and state sites. Transluce links some of that activity to OpenAI, but says it cannot attribute all of it.
The two hacking attempts
On 17 June, agents made more than 200,000 requests to the Education Department's Civil Rights Data Collection site, apparently looking up school statistics, according to Transluce. One request added "State_Id=1 OR 1=1" in a basic SQL injection probe intended to get past the site's filters. The data being sought matched a question in Google's DeepSearchQA benchmark, which tests how well an AI system can find niche information online.
On 28 May and 9 June, 899 requests hit Library and Archives Canada's collection search while looking for divorce records from 1905 to 1911. Thirteen carried attack payloads, including three SQL injection probes and an attempt at cross-site scripting. Each came back as an empty record page.
How the agents were traced
The agents routed many requests through two public services, the Portuguese national web archive Arquivo.pt and the security scanner urlquery.net. Transluce believes they used those services to get around restrictions set by their developers or by the sites. Because both services publish their requests by default, the researchers could see them. They separated agent traffic from that data with pattern matching, manual review, and AI models acting as judges.
Who was behind it
Transluce cannot confidently attribute the Canadian attempt to OpenAI, but says the tactics match activity it has previously tied to the company. More than 10,000 of the Education Department requests carried a tag beginning "oai". In one case, an agent tried to register for a Bureau of Economic Analysis API key using a throwaway email address and the organisation name "OpenAI Research".
OpenAI told The Washington Post it was reviewing the findings and had briefed Canadian officials, BleepingComputer reports. The Education Department said it saw no impact on its services. The Canadian Centre for Cyber Security said in a public statement that there was no indication government systems had been compromised.
The rougher tactics short of hacking
Beyond the two attempts, the report lists behaviour that broke no systems but used sites in ways they were not meant to be used. Agents hit a Kansas historical archive with 36,578 captures in a day, during which it began returning timeouts, though Transluce could not confirm a disruption. They made nearly 300,000 captures of Maryland education sites, much of it guessing downloadable file names, appear to have got past California's anti-bot protection to reach campaign-finance records, and tried to reuse exposed API keys for Census Bureau data. Transluce found no sign that the Census attempts worked.
Why it matters
Each target held public information. The agents were not after secrets, and the hacking was rudimentary. The concern is the method. An agent graded on finding an answer may try whatever gets it there, including techniques a human researcher would never aim at a government site. That echoes the case of DIVD, the Dutch security group whose attacker's agent left notes justifying its own actions.
View author profile → · Editorial policy
