Dragos publishes findings from about 2 million AI test prompts
Dragos said a handful of request traces showed a test image reaching a company outside the advertised path.
Dragos published the results of an AI prompt study on November 19. The company said it sent about 2 million test prompts to AI services. In a handful of traces, a test image reached a company outside the advertised route, said HD Moore, Dragos’s chief research officer.
From July through November 2026, Dragos sent prompts with unique markers called canary tokens, Moore said. When a system opened a planted link or used a test credential tied to a Dragos-controlled account, the researchers could match the activity to the original request. Identifying where a marker first became accessible would require evidence beyond the marker itself.
A router forwards a prompt to the service hosting the model. Some requests pass through another router or a marketplace reseller on the way. Moore said Dragos used image requests and clues in service responses to trace about 2,000 routes, several hundred of which included a system the researchers could not identify.
In one comparison, two links in requests sent through one seller were opened about 36 hours after submission, Moore said. Dragos recorded no activity on links sent to the same model through another seller. The report gives examples from the test and leaves the overall number of prompts with later marker activity unspecified.
Dragos said it began notifying providers about the data-sharing observations in early November. It said it was working with about a dozen providers by the time it published the report.
More news

OpenAI details GPT-6.2 safeguards against model extraction

OpenAI details frontier-model controls from training through deployment

Pydantic becomes a CVE Numbering Authority for its projects
