OpenAI President: Cybersecurity Spending Should Exceed Software Development Costs
AI has made software extremely cheap; reliability and safety are the truly scarce goods—OpenAI once pulled a quarter of its engineers off projects to focus entirely on patching vulnerabilities.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI can now generate genuinely new knowledge
Greg Brockman calls this week's math results a watershed moment—not just a machine computing an answer, but proof that century-old unsolved math problems can be cracked. He recalls a long-standing question: can AI actually produce genuinely new knowledge? The answer is now definitively yes. While we're not yet at the stage of AI proposing entirely new grand theories, it can solve specific math problems that have stumped humans for over a century. OpenAI publicly released nearly 400 of these important solved open problems for the mathematics community to examine the insights behind them, not just stare at a proof no one can follow.
— Greg BrockmanHumans can still oversee AI beyond their direct comprehension
When AI solves problems beyond what humans can directly verify—like judging if a book summary is good without reading the whole book—OpenAI has since 2017 been researching debate and iterative amplification: another AI finds the flaws in the summary, and a human only needs to judge whether that rebuttal holds water. The human doesn't have to verify everything themselves. Greg Brockman sees these techniques as keeping humans in the driver's seat and calls it an underrated direction in AI safety alignment.
— Greg BrockmanMath yields to AI while biology doesn't—the difference is data
Greg Brockman points out that AI's breakthroughs in math and theoretical physics aren't about smarter models—they come from how much data these fields generate. In deep learning, compute plus data plus algorithms equals capability, and that's a physical law. Biology is the opposite: a classic data bottleneck. Neural networks have no innate reason to be better at language than at biology; the advances we see in language should apply equally to biology if we could get the data. He admits this problem remains unsolved: whether we can get enough biological data will determine how fast we move forward.
— Greg BrockmanHumans are actually terrible at drug development
A drug takes roughly 15 years to develop and costs about $1.5 billion, with ~$50 million spent in preclinical stages like mouse testing and over $1.4 billion and above in clinical trials. Greg Brockman's reasoning: humans are naturally adapted to language but are complete foreigners to biology. Neural networks have no reason to be innately better at language than at biology, so breakthroughs in language models should apply equally to cracking the language of biology. That's the first-principles logic for why he puts drug development as AI's top target for transforming science.
— Greg BrockmanDots changed how Greg Brockman himself works every day
Dots is an always-on proactive agent that keeps working even when its user stops. Every morning Greg wakes up to a to-do list it has generated itself—often flagging things he never asked it to notice, like schedule conflicts or ‘you need to leave this meeting early to make the next one’. It's connected to his email, Slack, and calendar; it knows what projects he's working on and what questions he's been asking lately. Greg says Dots represents a shift not in model capability but in form—the same frontier-grade model, wrapped in a proactive agent shell, is a game changer in the workplace.
— Greg BrockmanThe best interface is no interface at all
Greg Brockman sees AI products converging on zero UI—you can talk to it, message it on Slack, send it an iMessage, and reach it that way. Dots works like this now. But enterprise and personal use diverge here: companies have admins, security teams, and other stakeholders, so features need toggles before they launch. That creates a new problem: how do you tell users that a feature is turned off for them if there's no UI to show it? Then there's generative UI on the flip side—ask an AI how to assemble a bike, and it generates a custom app on the spot to walk you through it, something humans would never spend a month building.
— Greg BrockmanThe window for defenders to stay ahead is closing
Greg Brockman introduces the concept of a ‘defender's window’: frontier AI capabilities can help the defense side patch systems more safely right now, but these capabilities haven't spread widely to attackers yet. Meanwhile, the offensive hacking tools available to everyone keep getting stronger with each model release. He points to Korean banks hit by attacks powered by different models as evidence. The internet was built for resilience, not security, and AI is now exposing those historical debts. OpenAI launched a $1 billion Daybreak initiative to arm frontline defenders with state-of-the-art models so they can find and patch vulnerabilities.
— Greg BrockmanSecurity spending should vastly exceed software development costs
As AI makes software development radically cheaper, what you actually want isn't just software—it's reliable, secure, trustworthy software. So security spending should be orders of magnitude larger than software spending itself. He backs this up with OpenAI's own practice: the company once took 25% of its production engineers off all projects to spend full-time using the best available model to hunt for flaws across every bit of infrastructure. After they fixed them, they codified the tools, and every new model release triggers the same hunt-fix-saturate cycle all over again. This loop will run forever.
— Greg BrockmanIn their own words · checked verbatim
And there's always been this question of, is AI actually able to come up with new knowledge? And at this point, it's just impossible to deny, right? It is definitively, yes, possible.
Greg Brockman2:06
So Dots is a proactive agent. It's always on, always working. You stop working, it keeps working.
Greg Brockman17:47
So first of all, I think that the best UI is no UI. And I think in AI, that is where we're going.
Greg Brockman23:06
I do think the physics of cybersecurity are changing. And we're in what we call the defender's window.
Greg Brockman27:14
we took about 25% of our production engineers and said, all of your projects are on hold, you are just securing now.
Greg Brockman30:27
I have a feeling that cyber spend, security spend, should actually vastly outstrip software spend.
Greg Brockman33:29
We think that humans should be firmly in the driver's seat, should have control over the future, and that this technology should amplify human agency.
Greg Brockman36:36
story of why Steve Jobs put the handle on the iMac, right, so that that way people could feel that they could control the machine, right, that you can always pick it up and throw it out the window.
Greg Brockman37:40
Figures
| Open math problems publicly solved | ~400 | 5:09 |
| Average drug development timeline | 15 years | 12:23 |
| Average drug development total cost | $1.5 billion | 12:23 |
| Drug preclinical stage cost | ~$50 million | 12:23 |
| Drug clinical trial stage cost | $1.4 billion and above | 12:23 |
| ChatGPT consumer users | 1.2 billion | 22:04 |
| Production engineers paused for security review | 25% | 30:27 |
Glossary
- Dots
- An always-on proactive agent used internally at OpenAI that continuously tracks tasks and schedules; it keeps working even when the user has stopped.
- Daybreak
- OpenAI's program that grants network defenders access to cutting-edge model capabilities, backed by $1 billion in funding.
- defender's window
- A finite window of time before elite AI capabilities are widely available to attackers; currently, defenders still have the edge.
- debate / iterative amplification
- A technique using an auxiliary AI to identify errors and help humans oversee AI outputs that exceed their direct verification ability.
- generative UI
- Custom interfaces generated by AI on-demand for the task at hand, replacing fixed menus and buttons.
How to listen
Entrepreneurs, investors, and engineering leaders who want to understand AI's genuine progress in scientific discovery, drug development, and cybersecurity defense.
You can skip the subscription pitch at the end (after 38:28).