Cloudflare CEO Matthew Prince said OpenAI and security researchers found a remote code execution (RCE) vulnerability in PHP and notified Cloudflare and other organizations about two months before public disclosure. PHP is used by about 70% of the web, and popular content management systems such as WordPress are built on it. The vulnerability let an attacker send requests to a server, take it over and run code on it—for example, read files and capture usernames and passwords. A compromised server could also be used for other criminal activity. Later, Wiz, a Google subsidiary, reported a second PHP vulnerability.

Prince said the Google and OpenAI teams worked with industry leaders, including Cloudflare, to address the issue. The effort also involved Automattic, the creator of WordPress; the PHP Foundation, which maintains PHP; and the U.S. government. Over two months, organizations coordinated their response to the risk. Cloudflare quietly enabled free protection for all its customers, including free-tier users. Prince described it as a temporary measure that bought time, not a substitute for installing the fix.

Prince said they saw an uptick and signs that information had leaked about 36 hours before the patch was released. After publication of the patch, attackers had enough detail to work out how to exploit the vulnerability: in the hours before its release, Cloudflare saw thousands of attacks across its network; afterward, it was seeing millions of attempted attacks per minute. Cloudflare’s protection helped, but Prince urged server operators to install the latest PHP version. The update is free and, he said, takes only a few minutes. Managed platforms such as WordPress often install updates automatically. He also relayed NSA guidance: if a patch cannot be installed, taking the server offline is safer than leaving it unprotected.

Prince called it difficult to decide who should receive advance access to information about new vulnerabilities, and when. The more people and organizations receive it, the greater the risk that it will reach someone who could cause real harm. At the same time, European banks and leaders ask why they do not get access to new technologies and say they feel vulnerable. Prince said that when Cloudflare was a small startup, it too asked why it was not told about new vulnerabilities; now Cloudflare receives such information, while startups ask how they can get access. He said there is no perfect answer. In his assessment, disclosure almost two months in advance gave organizations time to prepare and patch the issue, helping mitigate what could have been a disaster. He praised the OpenAI, Google and Wiz teams for their handling of the process, while acknowledging it was not perfect and calling it very good.

Matthew Prince described a problem involving root keys used in internet encryption: an Indian company had used the password “4” instead of the required minimum length of 328 bits, he said. He recalled that six to eight keys were affected and that the strongest password was “136.”

If an attacker can intercept network traffic, such certificates could let them impersonate Google or a bank. The industry distrusted the affected certificates in browsers and operating systems, but unupdated devices might continue to trust them. Prince advised people to install browser and operating-system updates.

In the interview, Matthew Prince distinguished finding vulnerabilities from exploiting them: current attempts to exploit the PHP flaw do not require AI, he said; attackers are using simple scripts to scan servers. But he believes AI is increasing how quickly an attacker can work out what to do after gaining access. Prince cited the compromise of a widely deployed Salesforce plugin by a Russian hacker about a year earlier. In his assessment, the attacker did not understand the Salesforce environment and acted inefficiently. AI could suggest what to look for and steal.

Prince said agent “swarms” remain expensive to run, potentially costing hundreds of thousands or even millions of dollars. As a result, typical cybercriminals are still using AI rather naively, though state actors such as China or Iran are a different case. His assessment is that AI currently helps defenders more than attackers by helping find vulnerabilities. He warned that the next two years will be fraught, as AI tools can patiently search the internet for problems. He expects frightening attack stories, but predicts that cooperation between frontier AI companies and organizations such as Cloudflare will ultimately make the internet more secure.

Discussing recent attacks on OpenCode, Railway, and an unnamed major inference provider, Matthew Prince suggested that attackers want more computing resources and access to models. Token theft and attacks on inference providers, he said, give them access to those tools. He said attackers have historically gone after valuable online targets and considers tokens and inference access especially attractive now. Prince advised keeping security patches up to date. He added that he did not know the details of the Railway attack and described a possible link to recent vulnerabilities as his hunch.

A Cloudflare representative said the company needs to move faster through the cycle of identifying threats and responding to them. Cloudflare uses AI tools and aims to build a “swarm of defenders” to counter automated attacks. He urged organizations to keep security patches up to date, but stressed that he does not know the details of the railway cyberattack; his suggestion that it may be tied to a recent vulnerability is only a hunch.

He said Cloudflare should develop partnerships with OpenAI, Anthropic and other companies to help defenders keep pace with attackers. He believes resources currently favor the “good guys” and says Cloudflare can identify threats in internet traffic, including attacks associated with North Korea and attacks involving swarms of agents.

The speaker expects major vulnerabilities to emerge almost every week on average over the next two years; he added that some substantial vulnerabilities expected to be disclosed in the coming weeks or months are already known to the company. This is his assessment and forecast, not a claim that all those vulnerabilities have already been publicly disclosed. In his view, in two years it will be nearly impossible to write code with classic errors such as division by zero, off-by-one errors and memory bugs, although business-logic errors will remain. He predicts that, despite several years of alarming headlines, the internet will ultimately become more secure thanks to AI tools.

Asked how computing resources might affect the geopolitics of cybersecurity, a Cloudflare representative said they could matter. He was nevertheless encouraged that defenders have more resources than attackers, which he believes will help them keep pace.

He stressed the importance of partnerships between frontier-model developers such as OpenAI and Anthropic and infrastructure and security companies, including Cloudflare. He said Cloudflare’s original thesis was to analyze large volumes of internet traffic with machine learning to identify attacks. He also said demand for Cloudflare services is growing and noted that the company wants people to be able to stay safe online without having to pay.

A Cloudflare representative predicted that, if current trends continue, automated traffic—including agent traffic—could be 1,000 times human traffic in five years. He noted that internet traffic doubled during COVID and that infrastructure nearly collapsed. In his example, a ChatGPT agent might inspect menus from 1,000 restaurants even though the user will choose only one, consuming resources across many sites.

In his view, the internet will need a new business model. One possible option is micropayments: an agent might pay one ten-thousandth of a cent to access a resource. The price could fall or become zero if the user agrees to reveal what the agent is looking for; in some cases, a site might instead pay for a visit. He also suggested that advertisers could target agents directly by matching their requests with relevant offers.

The speaker said Cloudflare is spending considerable time on the question and that the internet’s future business model will need to be worked out within the next five years, probably before the end of the decade. This is his forecast and hypothesis, not a claim that a micropayment system is already operating.

A Cloudflare representative voiced concern that fewer people are returning to original sources. He emphasized the importance of journalism and said there needs to be a system that compensates reporters for meeting people and gathering information. This was his view on supporting journalism as AI-generated content and answers may draw audiences away from original material.

After the presentation of a new aircraft, Trevor Milton responded to critics in the comments and said his company already had more than 50 orders or letters of intent from real buyers. The speakers clarified that the aircraft on display was a full-scale mock-up, not a flight-ready plane. Milton also claimed the company had secured nearly $1 billion in orders. These figures were presented in the discussion as his claims; no independent confirmation was provided.

Meta plans to issue non-voting Class C shares and distribute one such share for every existing Class A or Class B share at the start of next year. The hosts described this as effectively a 2-for-1 stock split: in their view, it would not raise money or change ownership percentages, but would allow Mark Zuckerberg to sell or transfer the economic portion of his stake while retaining control.

This is not the company’s first attempt to create Class C shares. In 2016, Facebook proposed the mechanism so Zuckerberg could sell or donate part of his economic stake without giving up control through his super-voting shares. After shareholders challenged the proposal and a Delaware trial was scheduled, the company withdrew it in September 2017.

In discussing the current plan, a host cited a New York Times account: according to that report, Zuckerberg had been negotiating with the board since July and effectively gave its members an ultimatum at a September 9 dinner—if they did not reach a deal, he would list his Palo Alto home for sale. The host suggested the dispute could involve both California’s proposed billionaire tax and Zuckerberg’s desire to retain control and stay close to his team; he also noted that Semafor offers a different account of the motives. The suggestion that selling the home was a threat because it could affect board members’ property prices was the host’s joking speculation, not the substance of the report.

As described on the show, Class B shares carry 10 votes per share, Class A shares carry one, and the proposed Class C shares would carry no votes. The host argued that the lack of voting rights might matter little to buyers focused on financial returns. He also expressed the view that Zuckerberg may have several reasons for seeking both to retain control and unlock the economic value of his stake. Potential tax consequences—including a possible 5% tax—and a move by Zuckerberg to Florida were discussed by the hosts as scenarios, not established outcomes.

The host called Meta’s willingness to invest capital in new technologies, including AI, rational, while noting that the returns on the company’s AI capital expenditure are debated. This was the participants’ assessment, not a confirmed investment outcome.

Kevin Williams, founder of Warp Sciences, said the company is building robotic laboratories in the US and using AI and robotics to help accelerate the discovery of new medicines. Williams said Warp is creating a preclinical contract research organization (CRO), not developing its own drugs: it is building infrastructure for other drug developers.

Williams gave two reasons for the effort. First, automation could enable more experiments, including around-the-clock lab operations, and generate large datasets for training AI models in science. Second, he said scientific and pharmaceutical capabilities need to be brought back to the US. In his assessment, substantial parts of drug-development programs take place in China; he likened this to the relocation of manufacturing abroad and warned that abandoning today’s manufacturing can leave a country locked out of future industries. He cited Andy Grove’s statement about that risk.

Williams described the optimization of a potential drug as a process that can take 3 to 5 years: researchers modify a molecule to improve its potency and solubility and reduce its toxicity. He said this stage is often outsourced to China and that Warp aims to bring such work back to the US using AI and robotics. He added that cost, rather than local opposition to laboratories, was the main factor behind moving this research abroad.

Williams said there is already a large market for US-based services: billions of dollars are currently going to overseas providers, and customers would prefer to conduct research in the US, where, he said, intellectual-property protections are stronger. He expects demand to grow, including as frontier AI labs and biotech startups become more involved.

The founder said Warp will combine purchasing existing robotics solutions with building its own. However, he stressed that not everything needed is available to buy, so the company is building from the ground up and vertically integrating. Its team includes chemists, robotics specialists, mechanical engineers, and AI researchers. This describes the company’s strategy and intentions; it does not establish that all laboratory capabilities are already operational.

The hosts asked whether biology would advance more slowly than software because experiments take longer. Williams replied that AI is already remaking software engineering and making progress in mathematics, but comparable mastery has not yet been achieved in the physical sciences, chemistry, and biology. He offered this as an explanation of the difference, not as a quantitative forecast of timelines.

In a discussion about applying AI to biology, chemistry and drug discovery, participants said progress in these fields has not yet matched what they have seen in software engineering. In their view, robotic labs matter both for increasing the volume of experiments and for generating data that can support the next generation of scientific AI models. They also connected a faster “model — experiment — data” loop with the possibility of faster progress.

One speaker suggested that shortening experiment cycles while applying models more often could have a multiplicative effect. Another said that if AI designs more drugs, more candidates will need to be tested, which will likely require physical infrastructure. Experiments in that infrastructure would also generate valuable data, the speaker said. They argued that labs should be built from scratch: in their view, existing CROs are difficult to refit after decades of operating differently, and many do not hire AI researchers or robotics specialists. The discussion also linked the need to develop robotic labs with reshoring this capacity from China.

Dylan Baxter, Plant Switch co-founder and CEO, said the company, founded in 2020, developed a material that feels and functions like plastic but biodegrades like plant matter. Baxter said it breaks down naturally without microplastics; he added that it could be discarded in the ocean or a backyard, although the company does not recommend doing so.

According to Baxter, Plant Switch set up a fully vertically integrated production site covering the process from raw material to finished goods. He described the company as the only one currently using this approach for a whole portfolio of products. The hosts also introduced Baxter as the founder of a company raising a Series B to expand Compost Zero into paper and coffee cups.

Stellius’s founder described the company as a developer of high-energy power systems in the United States, spanning directed-energy systems, high-power microwaves and compact fusion. The company intends to build more compact and powerful systems with vertically integrated U.S. production, including for military users. The founder said the main current application for high-power microwaves is defense: they can disable drone electronics, including in swarms, without expending expensive munitions. He also cited possible applications against the electronics of humanoid robots and satellites and in protecting space infrastructure; affecting people is not the objective.

Stellius is targeting a 1–2-mile range for complete area denial, with disruption and jamming at greater distances. The founder described the power range of the devices under development as 100 kilowatts to 1 megawatt. He said 100-kW-class devices can be deployed on mobile platforms, while megawatt-class systems are intended mainly for stationary use, protecting forward operating bases, data centers and borders. He stressed that microwave systems would be one component of a layered defense.

The company began around June and signed a $2 million contract to build pilot systems. It then raised a $5 million pre-seed round. The founder named Jamie Gall of Wave Function, Qubit Capital, Mana Ventures and Shervin Deshavar among the company’s supporters.

Separately, the company is only beginning research and design for a compact fusion system. The founder suggested that a successful development could change how things move, from hypersonic and supersonic vehicles to spaceflight; this is not an existing company technology but an objective and prediction.

Brian Manning of Zona Space Systems said the company launched six satellites earlier in the week to begin deploying its commercial satellite-navigation service, Pulsar. He described the service as bringing GPS into the modern technology era. That morning, the company announced a $250 million funding round at a $2.45 billion valuation.

The six satellites went up on a SpaceX Falcon rocket. The company is moving from rideshare launches, where it needs only a small fraction of a rocket, toward launching a full set of satellites. Manning said the satellites are being designed to work with any available rocket; the company is also preparing to fly on Starship when it becomes available to commercial customers.

The company expects to fit around 30 current-generation satellites on one rocket and 50–60 second-generation satellites. Manning said denser packing should roughly halve launch costs. Manufacturing is ramping up: he said the company is moving toward producing a couple of satellites per week. The company is applying automotive-industry lessons to production and aims to eliminate unnecessary requirements so it can build satellites more efficiently and cost-effectively.

Manning also explained that the navigation system operates as a one-way broadcast rather than a two-way connection. As a result, he said, once coverage is available in an area, the network can serve an unlimited number of users there without congestion caused by an increase in user numbers.

A Pulsar representative said the company’s system is designed to complement satellite communications with precise location and timing data where GPS performs unreliably. He cited GPS interference and jamming, everyday navigation errors, and the dependence of aviation, drones, time synchronization, and logistics on legacy systems. He said Pulsar designed its signal and network around user equipment so integration could require only a software update rather than a custom chipset. According to the representative, the company has demonstrated the system with devices ranging from chips for pet collars to military and aviation receivers, and has tested signal penetration indoors and performance against jamming. He said the stronger signal can be easier to receive and may reduce battery drain; that is the company’s claim, not an independent measurement.