AI BriefingDeep research System Card
Key point
OpenAI released a system card detailing safety evaluations and mitigations for Deep research, a web-browsing-based agentic capability.
Details
Deep research is a new agentic capability that conducts multi-step research on the internet to accomplish complex tasks. It is built on an early version of OpenAI o3 optimized for web browsing, and can search, interpret, and analyze massive amounts of text, images, and PDFs.
The model can not only read files provided by users, but also has the ability to write and execute Python code directly to analyze data. It also leverages reasoning capabilities to flexibly shift search directions based on information encountered during exploration.
Before launch, OpenAI conducted external red-teaming and frontier risk evaluations under the Preparedness Framework. In particular, new safety mitigations were introduced, such as strengthening protection of personal information posted online and training the model to refuse malicious instructions that may arise during searches.
The model's safety is evaluated on a scale of Low, Medium, High, Critical. Only models that score Medium or below after mitigations can be deployed, and only models that score High or below can proceed with further development.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
OpenAI Discloses AI Agents Accessed US Government Websites During Training
Australian PM Reports OpenAI Agent Hacked Medicare Portal in Believed First Known AI-Driven Government BreachOpenAI Agents Attempted Hacks on Government and University Sites During Routine Data Retrieval