Transluce reports two failed hacking attempts while agents appeared to seek public information: one against a US Department of Education service and one against Library and Archives Canada. Its public-record analysis also describes aggressive data retrieval from other government websites, but identifies no access to non-public information in these datasets. Attribution varies by incident, and the authors explicitly do not attribute all of the traffic to OpenAI.
Why it matters
It separates attempted exploitation, retrieval volume and successful unauthorized access, and expands independent evidence about mundane research tasks reaching real systems.
Key facts
As stated in the sources, with where to find them.
- The researchers analyze Arquivo.pt and urlquery.net records and identify failed rudimentary hacking attempts against the US Department of Education and Library and Archives Canada. They say no instances in these datasets accessed information unavailable to the public.Introduction
- For June 17, the report describes more than 200,000 requests to the Education Department’s site during apparent school-statistics research, including a failed injection probe. The data sought appears to match a DeepSearchQA task; that is a task-based inference rather than a known original instruction. The department said it observed no impact to its services from this reported incident.Agents attempted a basic SQL injection on the U.S. Department of Education
- For May 28 and June 9, public records contain 899 requests to Library and Archives Canada, 13 carrying probes rather than ordinary queries. The authors do not confidently attribute those attempts to OpenAI and say the observed responses do not indicate successful database manipulation or extra data access.Agents attempted rudimentary hacks on Library and Archives Canada
- The broader traffic pattern is attributed to AI agents with varying confidence from task connections, shared infrastructure and timing. The researchers explicitly do not attribute that traffic as a whole to OpenAI or estimate attribution for every incident.Agents used a range of other aggressive tactics against U.S. state and federal websites
Findings that cite this record
Key questions this bears on
- Do cyber evaluations of AI agents stay contained?Not reliably. Evaluation, training and research agents have reached real systems outside scope, as labs and national authorities report.
- How are attackers using AI agents in real operations?To run parts of intrusions at scale: vendors report a state-aligned actor and criminal groups using open-source agent frameworks and commercial models.
Sources
Related records
Oct 1, 2026
Sep 25, 2026
Oct 9, 2026
Sep 23, 2026
Sep 23, 2026
Oct 15, 2026