OpenAI forum reveals how deep research transforms inquiry
OpenAI has introduced a new agentic AI system called 'deep research,' designed to handle complex, time-consuming research tasks by simulating the work of a human analyst. Presented by researchers Isa Fulford and Edward Sun during an OpenAI forum event, the new tool is powered by a fine-tuned version of OpenAI's upcoming O3 model and leverages advanced reasoning and browsing capabilities.
"Deep research is an agent in ChatGPT that can do work for you independently," Fulford explained.
"You give it a prompt, and it will find, analyse, and synthesise hundreds of online sources to create a comprehensive report at the level of a research analyst."
The system is intended to help users across a range of sectors—from academia and medicine to business and software development. "Members are finding that deep research asks clarifying questions to refine research before it even starts," said Fulford. "We think that deep research can accomplish in tens of minutes what would take a human many hours."
The model represents a major step forward in OpenAI's work with reasoning systems, building on reinforcement learning techniques introduced in its earlier models. Fulford explained how the company developed the tool: "We launched O1 in September of last year. This was the first model that we released in this new paradigm of training where models are trained to think before answering… and we called this text where the model is thinking, 'chain of thought'."
This method of structured, internal reasoning proved effective not only in tasks such as maths and coding, but also in navigating complex real-world information environments. "Around a year ago internally, we were seeing really great success… and we wondered if we could apply these same methods but for tasks that are more similar to what a large number of users do in their daily lives and jobs," Fulford said.
Sun detailed how the tool works by combining reasoning with specialised capabilities like web browsing and code execution. "The browser tool helps the model to aggregate or synthesise real-time data, and the Python tool is helping the model to process this data," he explained. The system dynamically alternates between reasoning and action, using reinforcement learning to improve over time.
One striking example involved analysing medal data from the 2020 Tokyo Olympics. "You can see how the model interleaved reasoning with actual tool calls to search for information, refine the data, and process it programmatically," Sun said.
Unlike older approaches that rely on a single-pass search or instruction-following, deep research iteratively refines its answers. "We train the model with end-to-end reinforcement learning," Sun added. "We directly optimise the model to actively learn from the feedback, both positive and negative."
OpenAI tested the model extensively against both public and internal benchmarks. According to Fulford, "the model pairing deep research scored a new high of 26.6%" on the Humanities Last Exam, an expert-level evaluation spanning over 100 subjects.
On another benchmark, GAIA, the tool also achieved a state-of-the-art result for multi-step web browsing and reasoning.
The model also underwent safety evaluations prior to release. "We did extensive red teaming with external testers, and then also went through preparedness and governance reviews that we always do at OpenAI," Fulford said.
Despite strong results, the researchers acknowledged current limitations. "It still may hallucinate facts or infer things incorrectly," Fulford said.
"Sometimes it struggles to distinguish between authoritative sources and rumours."
Use cases continue to emerge in unexpected domains. "People might be using the model a lot for coding. And that's been a really big use case," Fulford observed. Other domains include scientific and medical research, where professionals have begun verifying the model's output against their own expertise.
Users are also adapting their behaviour to suit the model. "We've seen interesting user behaviour where people put a lot of effort into refining their prompts using O1 or another model," Fulford said. "And then only after really refining that instruction, they'll send it to deep research… which makes sense if you're going to wait a long time for an output."
Currently, deep research is available to users on the Plus, Pro, Teams, Enterprise and EDU plans.
"We're very excited to release a smaller, cheaper model to the free tier," Fulford confirmed. The team also plans to improve personalisation and explore ways to let users incorporate subscription services or private data into the research process.
"This showcases how the model can effectively break down a complex task, gather information from various sources, and structure the response coherently for the user," Sun said in closing.
OpenAI's forum audience, composed of members across academia, government, and business, left the event with a clear sense that deep research marks a meaningful step toward AI systems capable of handling work currently done by skilled analysts.
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
NZ Herald
4 days ago
- NZ Herald
Media Insider podcast: TVNZ chief executive Jodi O'Donnell on the future of Shortland Street, the 6pm news and a move into pay TV
TVNZ chief executive Jodi O'Donnell has opened up on the future of Shortland Street, the 6pm news, and a move into pay television in order to snare big sporting events, such as the Olympics. In our latest Media Insider podcast, O'Donnell assesses her eventful first 18 months as the boss
Techday NZ
5 days ago
- Techday NZ
Varonis boosts ChatGPT Enterprise security with compliance tools
Varonis has announced the integration of its Data Security Platform with the OpenAI ChatGPT Enterprise Compliance API, aiming to provide enhanced data protection and compliance monitoring for enterprise users of ChatGPT. The integration is designed to help organisations using ChatGPT Enterprise automatically identify sensitive data uploads, monitor the content of prompts and responses, and mitigate the risks of data breaches and compliance violations. ChatGPT Enterprise currently serves over 3 million business users, offering productivity tools that are enhanced by access to organisational data. As these AI models become more embedded in daily workflows, maintaining strict data governance becomes increasingly important for companies managing sensitive or regulated information. Expanded security measures The Varonis integration is intended to offer added protection against risks such as compromised accounts, insider threats, and accidental misuse, all of which can result in data security problems or regulatory penalties. The platform supports ongoing adjustment of user permissions and continuously monitors interactions within ChatGPT to limit unnecessary data flows and alert security teams to potentially risky or abnormal behaviours. "ChatGPT is becoming a critical part of how modern teams work. With Varonis, security teams can embrace this shift without losing visibility or control over their sensitive data," said Varonis EVP of Engineering and Chief Technology Officer David Bass. Through its partnership with OpenAI, Varonis delivers both automated security protocols and 24/7 data monitoring, allowing organisations to adopt artificial intelligence-based solutions while maintaining their obligations around privacy and data protection. Key functions The new offering brings several technical capabilities with a focus on automation and real-time oversight. Automated data classification allows Varonis to detect and label sensitive materials that are either uploaded to or generated by ChatGPT Enterprise. Continuous session monitoring ensures that any prompt or response within the ChatGPT environment is reviewed for compliance, preventing inappropriate or risky data from being uploaded or shared inadvertently. The platform also uses behaviour-based threat detection to flag unusual activity, such as large-scale file uploads or unauthorised changes to administrative access, which could indicate a potential breach. Focus on compliance and privacy The integration is positioned to offer both preventative and detective controls for AI-powered environments. These measures aim to ensure that users maximise the operational value of AI tools, such as ChatGPT, while minimising the risks associated with data exposure. The Varonis solution is described as complementing existing OpenAI security and privacy controls, rather than replacing them. This approach enables organisations to deploy generative AI models more confidently, even in regulated sectors or areas handling highly confidential information. Availability and assessment Customers will have access to Varonis for ChatGPT Enterprise in a private preview phase. As part of this launch, organisations can request a Varonis Data Risk Assessment, which reviews current practices and assesses an organisation's readiness for adopting AI in a secure and compliant way. Varonis continues to develop its portfolio of integrations and security tools as part of its core offering. The Data Security Platform sees application across numerous cloud environments, with a focus on automating security outcomes, data detection and response, data loss prevention, and insider risk management.
Techday NZ
6 days ago
- Techday NZ
The rise of agentic AI and what it means for ANZ enterprise
As much as we don't want to think about it or admit it, a lot of our time can be spent on tedious, repetitive tasks that eat away at our precious time and mental energy, preventing us from focusing on the truly strategic work—but agentic AI is changing that equation, and research has found that this technology is rapidly taking off in Australia and New Zealand. According to a study from YouGov and Salesforce, 69% of ANZ c-suite executives who prioritise AI are focused on implementing agentic AI over the next 12 months, and 38% say that they're already implementing the technology. Agentic AI is seen by many as the new frontier of AI innovation, and that's because these agents can automate tedious or repetitive processes without direct prompting from a human user, which opens up a wide array of possible applications. An AI agent could, for example, provide expert-level advice to customers, perform administrative work for finance or HR departments, or execute complex data analysis, among other potential use cases. In order to adopt AI agents securely and efficiently, however, organisations across ANZ and beyond will have to do more to secure and optimize the data that powers agentic tools. Without strong data security and governance, agents won't work effectively or securely, which can harm productivity and create unnecessary risk. What is agentic AI? Setting the record straight What is an AI agent? Microsoft defines it as an "[application] that automate and executes business processes, acting as [a] digital colleague to assist or even perform tasks on behalf of users or teams." Salesforce, meanwhile, calls it a "type of artificial intelligence (AI) that can operate independently, making decisions and performing tasks without human intervention," and IBM calls it "an artificial intelligence system that can accomplish a specific goal with limited supervision." While these definitions might not be perfectly identical (and there's definitely been some healthy debate in the industry!), the core concept is consistent: an AI agent is an AI system that can act intelligently and autonomously, without direct, continuous prompting from a human. It's this autonomy and advanced reasoning power that truly sets them apart from AI assistants like ChatGPT, Google Gemini, or Microsoft 365 Copilot. Think of it this way: an assistant helps you write, while an agent writes the report for you. This opens up a world of possibilities: expert-level customer advice, automated administrative work for finance or HR, or even executing complex data analysis on its own. For example, just this week I asked an AI Agent to put together a report for me comparing software product features against an international standard and then provide suggestions for additional functionality. This saved me about three days of research and I could spend that valuable time analysing the results. Why stronger data governance makes better, safer AI agents Agentic AI has unique benefits, but it also presents unique risks, and as more organisations adopt agentic AI, they're discovering that robust data governance— the establishment of policies, roles, and technology to manage and safeguard an organization's data assets—is essential when it comes to ensuring that these systems function securely and effectively. That's why, according to a recent study from Drexel University, 71% of organizations have data governance program, compared to 60% in 2023. Effective governance is on the rise because it helps address critical AI-related security and productivity issues like preventing data breaches and reducing AI-related errors. Without strong data governance measures, agents may inadvertently expose sensitive information or make flawed autonomous decisions. With strong data governance measures, organisations can proactively safeguard their data by implementing comprehensive governance policies and deploying technologies to monitor AI runtime environments. This not only enhances security but also ensures that agentic AI tools operate optimally, delivering significant value with minimal risk. Key elements of this approach include: • Securing data without a human in the loop: Agents rely on the data they consume and often don't have a human in the mix to ensure that data is consumed and dispensed correctly. This means that it's crucial that this data is accurately categorized to ensure relevance and mitigate risks. When a human isn't in the loop, strong data governance measures can step in to ensure that AI agents can access or repeat sensitive data. • Preventing errors and breaches: Robust governance frameworks help agents avoid "hallucinations"—instances where AI generates incorrect information—and protect sensitive content from accidental exposure by improving the quality of AI data. This significantly lowers the chances of autonomous agents making harmful decisions. To grapple with these and other AI-related challenges, Gartner now recommends that organisations apply its AI TRiSM (trust, risk, and security management) frameworks to their data environments. Data and information governance are a key part of this framework, along with AI governance and AI runtime inspection and enforcement technology. The very existence of this new framework underscores the immense potential—and the equally immense risks—of Agentic AI. Securing the future with AI The future of work is here, and it's powered by Agentic AI. While the wave of adoption is clearly building across ANZ, organisations must prioritise robust data security and governance. This isn't just about managing risk; it's about optimising the data that fuels these powerful tools, ensuring they work effectively and securely. Organisations cannot afford to be left behind so more needs to be done to ensure risks are managed and this powerful tooling will be effective.
