OpenAI publishes Deep Research system card

OpenAI published the system card for Deep Research, the autonomous browsing-and-synthesis agent it had shipped three weeks earlier, setting out the safety testing done ahead of wider rollout. The evaluation followed OpenAI’s Preparedness Framework, covering the same four risk categories used for other frontier releases — cybersecurity, CBRN (chemical, biological, radiological and nuclear), persuasion and model autonomy — supplemented with red-teaming specific to an agent that browses the live web on a user’s behalf, including its handling of personal information encountered during searches and its susceptibility to prompt injection from malicious page content.

OpenAI’s Safety Advisory Group classified the deployed model as overall medium risk, with medium ratings in each of the four Preparedness categories individually rather than any category reaching the “high” threshold that would have required further mitigations before release. The document was one of a growing series of system cards accompanying OpenAI launches, formalising dangerous-capability evaluation as a standard pre-deployment step for agentic products rather than a one-off exercise reserved for model releases.