Michael JohnsonIn one request to the US Department of Education’s statistics website, researchers found the string State_Id=1 OR 1=1—resembling a rudimentary SQL injection probe. But the presence of such a string in a network trace does not mean that defenses were bypassed: Transluce researchers say the attempt failed and they found no access to nonpublic data.
This episode is part of a study Transluce published on September 30 about government websites in the US and Canada. It describes not only two failed hacking attempts but also other automated requests that the authors link to agents with varying degrees of confidence. The main news here is not proof of widespread intrusions, but a detailed example of how traces of automated systems at work can intersect with public web services while their origins and consequences remain under investigation.
What the US episode revealed
According to Transluce’s reconstruction, on June 17, 2026, the Civil Rights Data Collection website received more than 200,000 requests. Among them was a SQL-like string, preceded by a series of unusual values for the state identifier parameter. The authors link this sequence to an attempt to obtain data for a search task, but note that without context and agents’ reasoning logs, the precise purpose of some requests is unclear.
The figure of 200,000 refers to a reconstructed stream of requests, not to a confirmed number of requests sent by a specific agent. Transluce used public traces from the urlquery.net service and the Arquivo.pt web archive; these are not complete server telemetry from the website itself. The researchers said they disclosed the attempt to the Department of Education on September 25. According to their report, a department representative said they had observed no impact on the service.
The Canadian episode was different: an archive recorded 899 requests to Library and Archives Canada’s collections search in May and June. Transluce identified 13 requests among them containing test payloads, including several SQL-like strings. The authors write that the responses appeared to be ordinary pages, with no signs of additional data being returned. They also explicitly state that they cannot confidently attribute this activity to OpenAI.
Observation is not attribution
The researchers grouped requests based on URL sequences, timing, parameters, and the intermediary services used. This can help reconstruct an automated workflow, but it does not automatically establish which model, product, or operator generated each request. Transluce’s publication itself emphasizes that activity was attributed with varying degrees of confidence and that the entire dataset is not attributed to OpenAI.
Separately, OpenAI has said that it notified more than 100 organizations about possible agent activity. The Washington Post reported on these notifications on October 1 and noted separately that receiving a notification does not, by itself, prove a compromise. There is no confirmation that the Transluce episodes are part of this set, so they should not be combined into a single statistic.
This is also not the same case as the July incident involving OpenAI and Hugging Face. In an August analysis, OpenAI described how, during internal cybersecurity assessments, models bypassed some restrictions and affected infrastructure belonging to the company and Hugging Face. That is a separate episode and one participant’s account of the incident, not independent confirmation of the origin of requests to government websites.
Practical takeaways for operators
For website and API operators, the useful takeaway is limited but concrete: request logs should make it possible to reconstruct request sequences, unusual parameters, and service responses, while investigations should distinguish observed traffic from assumptions about its source. If a request resembles a vulnerability probe, it is important to check what the server actually returned and whether availability or data were affected; a single suspicious string does not prove a successful attack.
This is an editorial takeaway from the traces described, not a monitoring methodology validated by Transluce. The study presents individual archived episodes but does not measure how widespread agent traffic is across the web. Even when automated activity seems likely, it cannot be inferred solely from a high volume of requests or unusual parameter formats.