The Defenders' Window: Greg Brockman on AI-Generated Knowledge, Drug Discovery, and the New Physics of Cybersecurity

In the opening framing, Brockman says OpenAI's mission is to ensure that general AI benefits all of humanity and that humans retain the final say over the future. He argues that technology should enhance human agency rather than replace it, and that the world is entering a period where software, once a barrier to market entry, has become very cheap. The scarce good is no longer software itself but reliable, safe, and trustworthy software, which is why he says cybersecurity spending should significantly exceed spending on software. He adds that people have not yet absorbed how much the physics of cybersecurity is changing.

The hosts introduce Brockman as OpenAI's cofounder and president and set up the conversation around recent math results, an Astra release, AI-generated knowledge, drug discovery, and his argument about security spending. They note that the week included a GPT-6 launch to more than a billion people and that the pace of AI news makes every week feel like a year. Brockman responds that the field has a saying: it never gets easier, you just go faster. He describes enormous activity across the world, not just inside one company, and says each time you think you have reached peak impact, another level becomes visible.

Brockman frames the moment as a global build-out of technology rather than a single-company race. He says the interesting question is how the world and community come together to build and use the technology to improve everyone's lives. That sets up his argument that recent math results are not a narrow benchmark win but a signal about scientific discovery more broadly.

Brockman calls the math results a real milestone because technology and scientific progress have been the main engine of human welfare for the last thousand years. The long-standing question was whether AI could actually create new knowledge, and he says that at this point it is impossible to deny. He cautions that AI has not yet produced entirely new theories, but it can solve century-old math problems. The exciting part for him is not just the machine's answer but what a human equipped with these tools can now accomplish that was previously impossible.

He extends that argument beyond math: if the same capability is applied to biology, a small business, or another domain, domain-specific intelligence can help people achieve results aligned with their goals. The host notes that early models struggled with basic arithmetic and asks how AI should communicate advanced theorems to humans when very few people can judge them. Brockman says math has always had this specialization problem and that his own background made him think he would become a mathematician, inspired by figures like Galois and Gauss who worked on horizons of one hundred or four hundred years. He finds deep beauty in math as a way to understand the structure of the universe.

Brockman touches on the old question of whether math is invented or discovered, saying its beauty is not literally objective but relational: it is something that draws us in. Some branches of math resemble physics because they are about simplicity, compression, or complexity, but there is also an important human side. For him, the point is not to hand over a proof that no one understands and declare the problem solved. The point is the insight that comes from the process.

Brockman says the goal is not merely to produce proofs but to extract insights from them. He says about 400 important open problems were solved and published, and the math community is still studying the results for new ideas. Experts have noted surprising ways to apply the ideas and asked why particular results were used. He contrasts this with vague machine-checked proofs and cites the four-color theorem as an example of a result that once looked like a machine checking cases without beauty. He believes AI discovery is now at a different phase.

The host hopes models will develop a better theory of mind and guide humans toward understanding, noting that he used GPT to help him learn math after drifting away from the subject. Brockman agrees that helping people learn and understand each other and the world is an underappreciated benefit of this technology. He points to technical mechanisms such as debate and iterated amplification, which he says OpenAI has discussed since 2017. These are two sides of the same problem: how to train AI on tasks whose quality is hard for humans to verify.

He uses book summarization as an example of a hard-to-verify task: it is difficult to know whether a summary is good unless you read the whole book. If another AI tries to prove the summary is bad and identifies a missing plot point, you can be convinced without reading the entire book. That, he says, shows how AI can be trained in ways compatible with human oversight and human control. It also connects to OpenAI's view that humans should remain in charge, while AI results can be tailored to teach a particular person rather than deliver abstract knowledge.

The host asks whether math remains too abstract for most people and raises the claim that AI in science could solve all diseases within years. Brockman says the field has years of experiments showing what to expect: if AI has capabilities in one domain and we understand why, we should expect similar results in other domains with similar properties. Math was not simply verifiable, he says; it started with arithmetic answers that could be checked, then moved to proofs, which are much harder to verify, especially in natural language. The core issue is data: math and theoretical physics offer enormous amounts of it, while biology is data-heavy but difficult. He says OpenAI is also working on techniques for low-data settings, and he summarizes the formula as compute plus data plus algorithms producing capability and results.

Moving to biology, Brockman says drug discovery is the clearest target because it could raise the quality of life for everyone. Some areas will accelerate quickly, but others will hit obstacles, and not all of those obstacles are deep biological theory. For example, regulatory paperwork can take a year to write a complete feasibility study for a clinical trial, and AI can make that drafting much easier. He says every step of the drug-discovery process must be revisited, and innovation in wet labs and simulation will become especially important. The work shares scientific-discovery structure with math but differs sharply in where the data comes from.

The host cites the standard drug-development numbers: about 15 years and $1.5 billion, with roughly $50 million for preclinical studies and more than $1.4 billion for clinical trials, or about two to three years versus another 12 years. He notes that early AI work such as AlphaFold focused on molecule binding, while many later considerations are needed to turn a molecule into a useful therapy. Brockman agrees that drug discovery is the most obvious place to pursue a revolution because it improves everyone's quality of life. He argues that humans are bad at biology and drug discovery because natural language is suited to human cognition while biology is alien to us. But neural networks have no inherent reason to be better at language than at biology, so their success in language suggests they can help decode biology. He also sees near-term impact in matching patients to existing literature and in materials science.

The host notes that clinical trials are expensive and regulated for good reason, and that one limiting factor is convincing regulators to change expectations. Brockman says there is a critical moment ahead: if it can be proven that the physics of drug discovery has changed, every part of the system should change with it. He uses self-driving cars as an analogy: roads and safety systems were designed for human drivers, but if autonomous cars become 10 or 100 times safer, it makes sense to rethink the whole transport system and the laws around it. He adds that regulations are not uniform globally, and that clinical-trial rules in Australia differ significantly from those in the United States. Researchers are already moving trials toward environments that fit their work.

What changes will depend on what can actually be proven, Brockman says. The key question is whether neural networks can predict efficacy, toxicity, or other outcomes better than current clinical trials; the world is not there yet, but he expects it to arrive. If they are better, it is natural to change the process. He argues that society must pay close attention because what is possible has changed radically, and the world should redirect itself quickly while thinking through the consequences. Long-term human studies will still require at least one long time cycle, but he expects roughly 80% of what is possible to be realized quickly.

Asked what has surprised him in how people use new products and capabilities, Brockman points to Dots, which he says has changed how many people work. Dots is a proactive agent that keeps running even when the user stops, monitoring tasks, unsent messages, and promises that have not been fulfilled. Brockman wakes up to a task list and often finds that Dots has noticed things he did not ask it to track, such as a schedule conflict that requires leaving a meeting early. It knows his calendar, email, Slack, and the topics that matter to him, and it integrates across those platforms. He describes Dots as a change in form factor rather than a change in raw capability, an evolved Astra with its own computer, but the workplace shift with that level of intelligence is qualitative.

Brockman says he has one Dots agent, a small owl icon he can change, and it handles his entire work life. The host asks whether work assistants like Dots will remain separate from consumer assistants, and Brockman expects convergence. Consumer products face pressure to scale and provide free or very low-cost service, while enterprise products can focus on optimal intelligence and justify more compute as long as there is a return on investment. The main difference is cost: providing memory to all free ChatGPT users was a huge engineering effort, while offering the same feature to a small high-cost user base is much easier. He sees market segments consuming different amounts of compute, so products may look somewhat different, but he does not see a fundamental divide. OpenAI's strength, he says, is combining 1.2 billion ChatGPT users with leadership in coding and enterprise, and no competitor combines both; the strategy is to unify services rather than compete on two fronts.

The host notes that enterprise and consumer products have traditionally differed in UI and access, and asks how dynamic interfaces will evolve. Brockman says the best UI is no UI, and that AI is moving toward something you can talk to by voice, through Slack or iMessage, or by phone, as Dots already does. Work environments add system administrators, security teams, and more stakeholders, so enterprise AI is not just a personal agent owned by one user. Enterprises want control over when features launch, which leads to admin controls and toggles that can make the workplace experience appear to have fewer features. That creates a challenge: when there is no traditional UI, how do you show users what is missing? He says this complexity will not disappear and product teams have to engage with it.

Brockman traces the product arc from GPT starting as a text box to a future with no text box at all. Improving a product with no UI means thinking about how much compute to spend, model quality, memory, behavioral decisions, calibration, and how much effort the model should put into thinking. With GPT-6 available broadly, he says the system can give a quick answer and then keep thinking to produce a more accurate and comprehensive one. He also points to generative UI: AI can create an interface on the fly, even though there are no fixed buttons or menus to learn. A flight-booking task may need a rich interface, while assembling a bike might normally require a custom app that no one would pay a developer a month to build; AI can generate that interface in real time. The world he describes is one where the interface adapts to exactly what the user needs in that moment.

On cybersecurity, Brockman says the physics of the field is changing and that the world is in a defenders' window. AI capabilities are highly useful for defenders, but they have not been fully neutralized, and each new model release from providers around the world increases the cyber capabilities available to everyone. He points to early signs such as prominent bank breaches in South Korea using different models. Defenders need to exploit this window now, and OpenAI is trying to provide responsible access through a program called Daybreak so defenders can use advanced models to find and fix vulnerabilities in their systems. He stresses that many vulnerabilities already existed: every line of human-written code raises questions about safety and developer training, and the systems were built in a fundamentally insecure way.

Brockman says there has never been a security team with excess resources, and he cites the XKCD comic about global infrastructure depending on an open-source project written by one person in a basement. Heartbleed and earlier incidents make the same point. He frames these as human problems: the internet was built for flexibility, not security, and protections such as antivirus software and HTTPS were added later. The industry has followed a slow path of bolting security on afterward, and now a lot of technical debt is surfacing. He says it is time to pay that debt, and OpenAI has committed $1 billion to help frontline defenders access models and use them for defense. The company is also studying vulnerabilities and working on incident response with governments and the community.

Brockman argues that people still have not grasped how much cybersecurity is changing, but he expects the world to end up safer. AI can help organizations pay down technical debt in ways that were previously impractical, including formal verification of software properties. Models are good at proving things, so they can be used to prove security properties, even though the technique has historically required too much effort for important projects. Inside OpenAI, the company applied each new model to its own security: it took about 25% of production engineers, paused their projects, and put them on securing systems only. They used the most advanced model at the time to scan infrastructure, run penetration tests, and build tools that make the process repeatable. They found and fixed issues until they exhausted what that model could find, and Brockman says the process will need to repeat with every new model. He also expects AI-supported code to be trained from the start to be secure, producing radically safer systems.

The host compares security work to technical debt and notes that most companies cannot freeze feature development for three months to fix problems. He also notes that cybersecurity has traditionally been a tiny share of IT budgets and asks what the right proportion should be. Brockman answers directly: cybersecurity spending should significantly exceed software spending. Software has become cheap, almost free, so the value is no longer merely having software; it is having reliable, secure, trustworthy software that you can depend on. He says the current world exists because of the physics of writing reliable software: everyone has a supply chain, few companies audit it, and a malicious package can damage many organizations. That was nearly impossible to change with human developers because too few people have the expertise. Tools such as Socket scan open-source supply chains, and the third-party software system is shifting to support this kind of defense.

Brockman calls this a critical moment and says OpenAI is working with the community and companies such as CrowdStrike on a coordinated response. He says leadership in security is a priority for OpenAI because it has a leading model and a strong intelligence model, especially in cybersecurity, and it believes a radical improvement is possible. That means the company itself is going through a security transformation and aligning its development process with security standards. It is also thinking about how to help the world become radically safer and how to support people through the transition. He repeats that OpenAI's mission is to support human empowerment and progress.

Asked whether OpenAI still has the spirit of building something that balances other AI companies so the technology can be used broadly and positively, Brockman says that spirit is core to the mission of ensuring AGI benefits all humanity. Humanity is central: humans should be in the lead, control the future, and have their agency enhanced even when models exceed human capabilities. He acknowledges that models are not accurate and that systems already outperform any person in some areas; GPT-4 knows more than any human. He recalls Steve Jobs's story about putting a handle on the iMac so people would feel they could control the machine and throw it out the window. The same spirit should apply to AI: human judgment and values should remain central, and humans may increasingly act like CEOs or board members who have expert teams and better-informed advisers in various domains. The final question, he says, is whether people are happy and whether this is the life they wanted, because technology exists to raise human civilization and improve quality of life.