Revolutionizing Cybersecurity and AI-Driven Science with OpenAI President Greg Brockman
Greg Brockman, OpenAI co-founder and president, discusses OpenAI's recent math results, in which roughly 400 open problems were resolved using AI agents, and why he believes AI can now generate genuinely new knowledge. The conversation covers AI oversight and communication techniques like debate and iterative amplification, the data bottlenecks that separate math from biology, the potential transformation of drug discovery and regulation, the proactive always-on Dots agent, the convergence of consumer and enterprise AI, why the best UI may be no UI, the cybersecurity 'defenders' window,' and why security spending should outstrip software spending while humans stay in control.
Mathematics as a milestone for machine discovery
Greg Brockman opens with the claim that technology has always been the fundamental driver of human quality of life, and that frontier models now let us massively accelerate scientific discovery. For him, the open question of whether AI can produce genuinely new knowledge is now settled: it is ‘definitively yes.’ He is careful to add that models are not yet producing entirely new theories, but they are resolving problems that have stood for a century or more. OpenAI has published roughly 400 important open problems that are now resolved, and the mathematical community is combing through them for insight. Domain experts report surprising ways of applying ideas and results they did not expect, rather than the incomprehensible proofs that have long troubled the field. He cites the Four Color Theorem as the counterexample, where a machine checked a large number of cases and the result had little beauty. He sees a new phase of machine discovery in which the value lies in the insights a proof reveals, not just in the proof itself.
Brockman comes from a math background and describes mathematics as partly human: its beauty and meaning are in the eye of the beholder, even though it connects to physical laws, simplicity, compression, and complexity theory. He also argues that the bottleneck in mathematics is not only producing results but communicating them. Very few people can adjudicate a complex theorem, so the real goal is to surface the underlying insights in a form humans can absorb. He is optimistic that as models improve at modeling what humans know, they will also become better at guiding people toward new understanding. He personally illustrates this with his own experience relearning math with AI tools.
Communication, oversight, and verification
This theme connects to AI safety. Brockman points to debate and iterative amplification, techniques OpenAI has discussed since 2017, as ways to keep humans in the driver’s seat when AI solves problems too hard for humans to check directly. His example is summarizing a book: judging whether a summary is good is hard unless you have read the whole book. If a second AI can find a passage that contradicts the summary, a human can evaluate that single argument without reading everything. In his view, training AI this way serves oversight and also makes outputs more useful to people, tuned for human understanding rather than abstract knowledge.
Why math moved first and what transfers to biology
When asked why mathematics has advanced so quickly, Brockman distinguishes several factors. Models have reached a certain raw capability level. Math problems were originally easy to check, such as arithmetic with a numeric answer, but natural-language proofs are much harder to verify, though formalized proofs are checkable. The core insight he emphasizes is data. Mathematics and theoretical physics produce lots of data, while biology is a severe data bottleneck. He summarizes the deep learning formula as compute, data, and algorithms yielding capability and results. OpenAI has developed techniques to work with less data, and it is already applying these methods in biology. Some areas will accelerate quickly, but others depend on whether the data can be obtained, which will require wet labs and better simulation.
Drug discovery and the full pipeline
Brockman argues that drug discovery is the clearest and most important application to transform. He describes the pipeline as extremely long, roughly 15 years and about $1.5 billion, with a large share spent on clinical trials. Early work such as AlphaFold focused on molecule design and binding, but the bottlenecks are often downstream. Regulatory filings alone can take a year. He notes that humans are poor at biology, whereas language is what humans are naturally tuned for. Since neural nets are not obviously better suited to language than to biology, the same first-principles reasoning that produced results in natural language should be able to crack the language of biology. He expects short-term impact in parallelizing many steps of the pipeline and in finding existing papers that contain cures for particular conditions, and he sees similar opportunities in materials science.
Regulation and changing the rules
The conversation turns to regulation, which is a major constraint on drug development. Brockman expects the main obstacle to be convincing regulators to change expectations around approvals, but only if the physics of drug discovery have demonstrably changed. He draws an analogy to self-driving cars: if they become ten or a hundred times safer than human drivers, society should rethink the whole transportation system, including where autonomous vehicles are permitted. He notes that clinical trials already move across jurisdictions to more favorable environments, and that the exact changes depend on what can actually be proven, such as whether a neural net can predict efficacy or long-term outcomes better than a trial. He expects around 80% of what is possible to arrive quickly once proven, while acknowledging that long-term outcomes are hard to validate.
Dots and the proactive agent
Brockman describes Dots as a proactive, always-on agent that keeps working even when he stops. Each morning he finds a to-do list it has assembled from messages, commitments, calendar conflicts, Slack, and his inbox. It notices things he never asked it to notice and understands his projects and current focus. He emphasizes that Dots changes the form factor rather than the underlying capability, since it is an Astro-level model with its own cloud computer. In a workplace, that form factor is a game changer. He runs a single Dots instance that covers his entire work life.
Consumer and enterprise convergence
On whether work assistants like Dots will diverge from consumer assistants, Brockman expects convergence. The main difference he sees is cost. A mass-market free product must scale cheaply, while a busy professional will pay for more compute if the returns justify it. OpenAI has invested heavily in giving free ChatGPT users features like memory. Enterprise and consumer products may differ in functionality and in the controls needed for admins, but there is no fundamental difference in the underlying AI doing things on someone’s behalf. OpenAI’s advantage is serving over a billion ChatGPT consumers while also running a leading coding and enterprise business, which lets it unify these efforts into one competitive front. He also reports that people are quitting jobs to start companies because they rely on AI to handle the barriers to entry.
The future of UI
Brockman argues that the best UI is no UI. The eventual interface is simply something you talk to, through Slack, iMessage, or any surface. Enterprise settings differ because companies need admins, security teams, and control over when features roll out, which creates the challenge of surfacing missing functionality to users when there is no fixed UI. He sees the evolution from a text box in ChatGPT to a world where the AI molds itself to you. Models can now think longer before answering, returning a quick answer first and a more complete one later. Generative UI lets the AI create bespoke interfaces on the fly, such as a flight-booking interface or an app to assemble a bicycle, which would previously have required someone to write a one-off app.
Cybersecurity in the defenders’ window
Brockman believes the physics of cybersecurity are changing and that we are in a ‘defenders’ window.’ Peak frontier capabilities can help defenders secure systems, but these capabilities are not yet broadly diffused. Broadly available cyber capabilities are rising with each model release across providers, and early impacts are already visible, including high-profile bank hacks in South Korea. Defenders must use this time. OpenAI’s Daybreak program provides responsible access to advanced models so defenders can find and fix vulnerabilities. He frames the underlying problem as human: the internet was built for resilience rather than security, and security was bolted on later. Flaws such as Heartbleed and the XKCD example of global infrastructure depending on one unpaid open source project are symptoms of accumulated tech debt that now comes to the surface. OpenAI committed a billion dollars to help frontline defenders access its models. Internally, it pulled about 25% of production engineers off projects to secure systems using its peak model, then repeated the process until the model’s findings were saturated. Future models will require repeating this, but AI-assisted code can be trained to be secure from the start. Formal verification of software security properties, which models are good at, could become practical at scale.
Security spending and the software supply chain
Brockman and the host discuss a likely transitional period with more attacks before defense catches up. Brockman is optimistic that defenders can outstrip attackers, but this requires coordinated responses, investment in critical infrastructure, and resources for organizations that have historically lacked them. He argues that cyber and security spending should vastly outstrip software spending. As software becomes cheap enough to be the barrier to entry, the valuable thing is software that is reliable, secure, and trustworthy, and resilience becomes the core value. Supply chain security is a key example. Companies rarely audit their dependencies, and malicious packages have compromised many organizations. AI makes it possible to audit at a scale human developers could not manage. Third-party companies such as Socket scan open source supply chains, and OpenAI partners with companies like CrowdStrike. OpenAI views leading on security as a core priority given its frontier models.
Human agency as the guiding principle
The closing theme is OpenAI’s human-centric ethos. Brockman restates the mission: ensure AGI benefits all of humanity, keep humans in the driver’s seat, and use technology to amplify human agency, even as models surpass humans in certain capabilities. He notes that models are jagged, exceeding people in some dimensions while failing in others. He invokes the story of Steve Jobs putting a handle on the iMac so people feel they control the machine, and argues AI should work the same way. Humans may become more like CEOs or board members, relying on domain experts and teams that may have better ideas, while keeping judgment and values. The guiding question is whether this makes people’s lives better. Technology should lift up civilization, and that, in his view, is what we should seek from AI.