ExpertsFMThe Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman
No Priors: Artificial Intelligence | Technology | Startups · October 15, 2026
5 Selected Lessons:
case · 2 heuristic · 1 reasoning_pattern · 1 framework · 1
The defenders window and daybreak program
Source
No Priors: Artificial Intelligence | Technology | Startups — The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman (October 15, 2026) · at 27:02 ▸
When this applies
when a powerful new capability is temporarily asymmetric — available to responsible actors before it becomes commoditized and available to bad actors too
The takeaway
When you hold a temporary capability advantage, use it to proactively harden your position rather than treat the advantage as permanent — the window will close as the capability diffuses.
Why it works
Because broadly diffused attacker capability ratchets upward with every public model release, there's a finite window where defenders with privileged access to peak capability can find and fix vulnerabilities before attackers gain equivalent tools.
One level up
Treat any capability asymmetry as a countdown, not a steady state, and front-load defensive investment accordingly.
Where it stops holding
This stops working once attacker-accessible models catch up to defender-accessible peak models, closing the asymmetric window; the strategy also fails if defenders don't act proactively during the window.
Detail
procedure:
- Recognize that frontier AI capability on offense/defense is not yet broadly diffused — a temporary window exists.
- Give responsible defenders privileged access to peak capability models (daybreak program) ahead of general availability.
- Have defenders proactively hunt for and fix vulnerabilities in their own systems during this window.
- Expect the window to close as broadly diffused capability (via model releases from all providers) catches up.
rationale: High-profile hacks (e.g., South Korean banks) show attacker capability is already rising, making urgency real and the window finite.
In their words
We're in what we call the defenders window. And the way that I think about this window is that we have peak frontier AI capabilities that can help defenders secure themselves. They're not broadly diffused yet, but that the broadly diffused capabilities are ratcheting.
Regulation follows proven superiority, not sentiment
Source
No Priors: Artificial Intelligence | Technology | Startups — The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman (October 15, 2026) · at 15:02 ▸
When this applies
when a new technology outperforms a legacy process that is protected by regulation or institutional inertia
The takeaway
Don't try to change rules through argument alone — first generate incontrovertible comparative proof that the new approach outperforms the old one on the metric the rule was designed to protect.
Why it works
Regulators and institutions will only rationally update rules once there's unambiguous proof the new approach delivers better outcomes (e.g., safety, efficacy) than the old one; proof, not advocacy, is what unlocks system-wide change.
One level up
Structural change follows demonstrated superiority; design your roadmap to produce that proof before expecting institutional change.
Where it stops holding
This stops working when the proof itself is contested (e.g., short-term data can't predict long-term outcomes) or when regulators are not solely outcome-driven.
Detail
approach: Identify the metric regulation is meant to protect (safety, efficacy); build the technology to clearly outperform the legacy baseline on that metric; then expect/push for regulatory reorientation.
signals:
- Self-driving cars reaching 10x-100x safety vs. human drivers as the trigger for rethinking transportation regulation
- Neural net predictions beating actual clinical trial predictive power as the trigger for rethinking trial requirements
rationale: If AI can prove it delivers unambiguously better results than the current regulated process, the system should and likely will reorient around it — though verifying long-term outcomes remains a hard gap.
In their words
I think there is a moment here where, and there's more work to be done to reach this point, but if you can really prove that the physics of drug discovery have changed, then of course it should trickle through to every part of the system.
Capability transfers where checkability and data exist
Source
No Priors: Artificial Intelligence | Technology | Startups — The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman (October 15, 2026) · at 8:42 ▸
When this applies
when deciding whether a capability that worked in one domain will transfer to another domain
The takeaway
Before assuming a capability will generalize to a new domain, check whether the domain shares the same data and verifiability properties that made the original success possible.
Why it works
Deep learning capability is driven by compute, data, and algorithms; domains with abundant, checkable data (math, physics) let models be trained and verified at scale, while domains with scarce data (biology) bottleneck progress regardless of model intelligence.
One level up
Explain your wins mechanistically before predicting where else they will repeat.
Where it stops holding
This stops working when a domain lacks any proxy for checkability or when data truly cannot be generated or substituted via synthetic means.
Detail
rule: AI capability advances fastest where there is abundant, checkable data; it stalls where data is scarce, even if raw model intelligence is high.
threshold: Mathematics and theoretical physics have abundant checkable data; biology is 'very data-bottlenecked' per Brockman.
rationale: The fundamental physics of deep learning is compute + data + algorithms = capability; without data, intelligence alone does not solve the problem.
In their words
But the thing that actually I think is the core insight is really about data, right? It's that you can get lots of data in mathematics, you can get lots of data in theoretical physics. Can you get lots of data in biology? That's obviously something very data-bottom-length, very difficult.
Brockman's description of DOTS as a proactive, always-on agent
Source
No Priors: Artificial Intelligence | Technology | Startups — The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman (October 15, 2026) · at 18:24 ▸
When this applies
when designing any automated system meant to reduce a user's cognitive load around recurring, scattered obligations
The takeaway
In Brockman's view, a high-value form factor for an assistant is proactive surfacing of overlooked obligations, not just responsive answering of explicit queries.
Why it works
By continuously running in the background and integrating across multiple data sources (inbox, chat, calendar, stated commitments), the agent, as Brockman describes it, surfaces unprompted, contextually relevant information instead of waiting to be asked, which in his account narrows the gap between 'what the user should know' and 'what the user remembers to check.'
One level up
Shifting from a reactive to a proactive interaction model changes the capability-per-interaction ratio more than raw intelligence gains do.
Where it stops holding
This stops working when the user's context (calendar, messages, commitments) isn't integrated or accessible to the agent, or when false-positive alerts erode trust.
Detail
situation: Brockman, as a busy executive, needed to track to-dos, scheduling conflicts, and context across many data streams (Slack, inbox, calendar).
decision: OpenAI built DOTS as an agent designed to run continuously in the background, integrating across tools and surfacing unasked-for insights.
outcome: Brockman reports it 'has really changed how many of us work,' describing it as surfacing missed to-dos and calendar conflicts each morning without being asked.
lesson: Brockman's experience suggests a proactive, contextually integrated agent form factor can deliver outsized value even without new underlying model capability — a change in form factor, not capability.
In their words
I think DOTS has really changed how many of us work. And the thing that DOTS represents is dots is a proactive agent. It's always on, always working. You stop working, it keeps working.
Internal 25% engineering freeze to saturate security vulnerabilities per model release
Source
No Priors: Artificial Intelligence | Technology | Startups — The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman (October 15, 2026) · at 29:52 ▸
When this applies
when a powerful new tool becomes available that could expose or fix a backlog of neglected structural risk (tech debt, security debt)
The takeaway
When a capability jump occurs, dedicate a fixed, substantial resource allocation to apply it exhaustively against your own weaknesses before returning to normal operations, and repeat this process with each future capability jump.
Why it works
Temporarily reallocating a meaningful fraction of engineering capacity to use the newest, most capable model against your own systems surfaces vulnerabilities that previously went undiscovered because manual auditing was too expensive or slow, and running this to 'saturation' ensures the model's current capability is fully exploited before stopping.
One level up
Institutionalize a repeatable response ritual for capability jumps rather than treating each one as a one-off event.
Where it stops holding
This stops working when the model's current capability is insufficient to find deeper classes of vulnerabilities, or when an organization can't afford to pause a quarter of production work even temporarily.
Detail
situation: OpenAI needed to proactively secure its own infrastructure as new, more capable models became available.
decision: Put ~25% of production engineers' other projects on hold; task them with applying the peak intelligence model to probe all infrastructure for vulnerabilities and build reusable tooling to repeat the process.
outcome: The team found and fixed a number of issues and 'saturated' the issues findable by that model generation; Brockman expects this to repeat with each new model.
lesson: A repeatable, resourced ritual of 'full-capability audit to saturation' after each model upgrade can systematically pay down accumulated security debt over time.
In their words
So, for example, we took about 25% of our production engineers and said, all of your projects are on hold. You are just securing now. And took our peak intelligence model at the time, had it poke at all of our infrastructure, raising a bunch of things, built some tools so that it's easy to do it again over time.