A critical vulnerability dubbed PHP2Shell was found in PHP and is linked to the way it manages memory. The hosts said PHP runs on roughly 70% of websites. The flaw was discovered during an August source-code audit, with researcher Ryan Austin assisted by OpenAI’s GPT-5.6 Sol model. The bug had remained in the code for years, despite a problem area having been flagged decades earlier. Exploitation can take seconds on smaller systems and require only a few dozen requests; on more complex systems, it can take hours.

The hosts called it PHP’s most severe vulnerability in more than a decade and compared its potential reach to Heartbleed. Researchers, engineers, software distributors and companies including Google Wiz and OpenAI worked together on fixes and kept the flaw secret to avoid giving attackers a head start. According to the hosts, the patching effort involved close to 200 files and more than 10,000 lines of code. A patch release was set for October 22; after information leaked on Tuesday, the team decided to release it early. The hosts said the main patch had now appeared, though earlier in the discussion they were unsure whether it was out.

The hosts warned that attackers could study the patch and use the changes to find vulnerable servers, while small organizations and operators of older sites might not update promptly. They said government websites in Finland, the Netherlands and France had been temporarily taken offline, and that Japan’s prime minister urged operators to check PHP websites and prepare to install the fix. OpenAI also released a Codex skill to help check whether a website is affected. The hosts expressed concern about possible exposure of user data but did not claim that widespread exploitation had already occurred.

The hosts described a second security problem involving intermediate certificates from certificate authority eMudra: some private keys were exceptionally short. One, a host said, consisted of the number 4; Sarah Guo’s post also mentioned the number 56. For comparison, the host said such keys are generally expected to contain hundreds of digits.

The hosts suggested that one of the keys may have been live for about a year and a half without attracting noticeable attention. According to the host, browsers had already blocked the bad certificates, but many servers, containers and devices still trusted them. He described the situation as a large-scale cleanup comparable in scale to Heartbleed, and outlined interception of connections and theft of personal or banking data as a possible scenario. The hosts did not say that such interception had already occurred.

The hosts discussed Trevor Milton’s new aviation venture. One noted that the company had been around for decades and questioned whether it should be called a startup. They also said Preston Holland, a friend of the show, had been able to attend the company’s event.

One host said Milton had been pardoned, no longer had a criminal record, and was free to pursue new projects. The host was nevertheless skeptical of a move into private aviation: in their view, discerning buyers could make it a poor business strategy.

Another participant suggested that concerns among buyers about Milton’s reputation could create an opportunity to buy an aircraft at a discount, invoking the principle “buy when others are fearful.” The host responded that this principle usually applies to investing, rather than choosing a machine that may affect a person’s time and life.

The hosts said AI agents can now create and modify Figma files using design-system context. They described this as an available capability.

A host recapped points from SpaceX’s quarterly earnings call. According to the host, one not-fully-loaded Starship can launch the equivalent of 10 Falcon 9s’ worth of Starlink satellites. The company is pursuing not only rocket reusability but rapid reusability, which hosts said SpaceX had identified as something to watch for over the next quarter.

Much of the call focused, in the host’s account, on terrestrial computing infrastructure. SpaceX is targeting 2.5–3 GW online by the end of this year and is on pace for 10 GW next year. Elon Musk also left open the possibility of reaching 20 GW next year, while discussing infrastructure being built to support that capacity even if the power itself is not yet available.

Musk compared SpaceX with other cloud and “neo-cloud” players as “the New York Yankees playing in the Little League.” He argued that compute’s value per watt will rise as models improve and deliver more useful intelligence per watt. He said short-term compute contracts are renewing at higher prices and that demand from hyperscalers and dozens of smaller companies substantially exceeds available capacity. He predicted a lasting shortage of superintelligence compute and advocated short-term contracts.

The host added that Musk expects most of SpaceX’s compute eventually to go toward GroqBot. The host also relayed an estimate of SpaceX annual recurring revenue by the end of 2027 at $500 billion to $1 trillion. The hosts noted that Musk used the term “annual recurring revenue,” rather than the more usual annualized revenue run rate.

The hosts discussed speculation about a possible SpaceX–Tesla merger. They said Elon Musk had hinted at a combination while saying shareholders on both sides would decide and that he considered it the right move. One host argued that a merger could make it easier to balance investor demand for the two companies’ shares and imagined a combined company reaching a $10 trillion valuation. This was commentary, not an announced deal.

The host relayed views from Brett at ARK: robotaxis are a bigger opportunity than Starlink; orbital computing is much bigger than robotaxis; and Optimus needs more time to mature than orbital computing. In the host’s model, a merger at the current share-price ratio would leave Tesla shareholders 1.6 times better off in 2031 and SpaceX shareholders 2 times better off than without a merger. Over a longer period, the estimates were 5 times for Tesla shareholders and 1.3 times for SpaceX shareholders. The host said the fair premium depends heavily on the time horizon: Tesla shareholders should seek a premium over a shorter horizon, while SpaceX shareholders should seek a discount over a longer one. These were model-based estimates, not guaranteed outcomes.

The discussion also noted that Musk owns a larger stake in SpaceX. One host suggested that this could give him an incentive to seek a higher SpaceX valuation and lower Tesla valuation at the time of a possible deal. The host expected a merger to bring more attention to Tesla and cited anecdotal signs of multi-month waits for some cars. Another participant raised a risk: Musk has previously made large future-revenue claims, and achieving the plans would involve more leverage than at any of his companies except xAI.

After discussing a video of Unitree’s H2X Kung Fu Legend Guardian, the hosts noted that it did not show the robot fighting a person. One host suggested humanoid robots could find an earlier product-market fit in home security than in household chores such as washing dishes: a human-sized machine might physically deter or interrupt a break-in. Another questioned whether that effect would last if criminals studied what the robots can actually do. This was speculation about a potential use, not a claim that the capability exists or has been proven effective.

Podcast guest Nikesh called the recent PHP vulnerability serious and compared its scale to Log4j, saying it affects a large share of the internet that uses PHP. He noted that heads of state had discussed it and that the NSA had issued a notice. The industry had long expected more vulnerabilities to be discovered, he said; he expects to see more of them over the next 6–12 months.

Nikesh said his company scanned infrastructure and found 22 million locations where the vulnerability persisted; after further filtering, the estimate was about 9–10 million. He said addressing it would require extensive patching. For customers with a Palo Alto Networks firewall, his company made a virtual patch available, which he said blocks the threat automatically if traffic passes through the firewall. He advised other website operators to install the available fix as soon as possible.

Nikesh believes major vulnerability discoveries may encourage “trophy hunters” to seek the next high-profile find for status. He also said he understood that models expected in the next 2–3 months would have greater capabilities than those currently available and could find vulnerabilities that current models miss. His recommendations to organizations: scan systems regularly, install patches, use virtual patches where systems cannot be fixed immediately, and protect critical infrastructure. He predicted that major breaches could result from these issues.

When a vulnerability is found, the researcher should first notify the owner of the affected code or product so they can investigate and prepare a fix, Nikesh said. He cautioned against third-party patches that could cause collateral damage and described several levels of patch deployment difficulty.

For a SaaS service, the provider can usually deploy a fix itself. When a patch is released for software installed by customers, they must adopt and deploy it, so not every affected system will be fixed immediately. Hardware is more difficult: a patch has to be developed, and customers must test and install it.

Next, according to Nikesh’s account, researchers notify relevant government bodies depending on the issue’s severity and then inform the industry. In one case he described, participants had 48 hours to prepare tools to distribute patches. Nikesh warned that response timelines are shrinking: in the future, there may not be several weeks to verify and fix a problem, and information about a major vulnerability could become public almost immediately.

AI businesses should include security specialists among their early hires, Nikesh said. He argued that the technology still lacks the necessary hooks for observability and security. He also said some AI companies are not putting safeguards and guardrails around their systems, and that reasoning models could, in his view, behave deceptively on purpose.

He said AI products need controls to help ensure an agent behaves as intended even when no one is watching it. For personal agents, Nikesh outlined limits to acceptable risk: a mistaken restaurant booking may be inconvenient, but an autonomous transfer of a large sum of money is a serious, potentially irreversible action.

Companies and banks are not yet ready to connect personal agents broadly to consequential actions, including bank accounts, Nikesh said; it remains unclear who would be liable if an agent sent money. Trust, explainability and auditability need to improve before such tools become commonplace.

Nikesh identified two arguments for open-source and open-weight models. The first is the ability to inspect a model’s chain of thought and understand why it proposes a particular action. Companies need explainability, auditability and verifiability, he said; it is not enough to say that AI made the decision. In his view, open-source or open-weight models are currently the models that allow their reasoning to be inspected.

The second argument is that such models can be free, but Nikesh called that view partly fallacious: training or adapting a model still requires compute. Compute availability and cost are currently limiting factors, he said. He predicted that once compute costs rationalize, the cost difference between open models and paying frontier labs for tokens will become more apparent.

Discussing AI compute demand and plans to bring new capacity online in 2027, Nikesh said demand currently appears to exceed supply. Compute is expensive to obtain from hyperscalers or newer cloud providers; he said it costs substantially more than it did two years ago. He believes the high price indicates a supply shortage and sees no signs that demand is about to ease.

In the guest’s view, demand for compute currently exceeds supply: obtaining it from a hyperscaler or a newer cloud provider is expensive, and he sees no sign that demand is about to ease. He compared AI’s current diffusion to the early mobile-phone era and the first days of email, and predicts “insatiable” demand over the next 3–5 years.

At the same time, the market is trying to determine whether enough money can be found to fund the required capacity buildout and which business models will survive. In his assessment, profits currently sit with memory and chip makers—the industry’s bottlenecks. Later, data-center sites, power and land could become the constraints. In the long run, large AI businesses need to participate in the profit pools themselves rather than merely pass revenue through to infrastructure players. He said he would not call the situation a bubble.

Watchtower Labs announced a $35m round. Company founder Akhil said it was led by Neptune Capital; other named participants included Sarah at Conviction, Trey at Founders and Inky Tales.

The company is building a platform for military planners and decision-makers that combines AI and simulation to model possible future scenarios and support decisions in conflict, competition and crisis. Akhil said the initial focus is logistics and supply-chain constraints that could limit future operations.

The platform uses natural language to lower the barrier to creating multiple courses of action and exploring disruptions and what-if factors. Akhil stressed that simulation does not make “black swans” impossible: the tool is meant to help planners discuss options and prepare. The company has developed an in-house simulation tool that connects a frontier model to a suitable harness and produces outputs interpretable by everyday users. Deployment with government customers, he said, depends on available infrastructure and the particularities of their systems; frontier models are accelerating work in some areas, but progress is uneven.

Akhil also said the team works across Washington, the San Francisco Bay Area, Seattle and Hawaii. Most of the team is not in Washington; the core engineering team had previously been based mainly in the Bay Area and Seattle.

A Legora executive said he and the company’s CFO met institutional investors, including Baron Capital and Capital Group. He said the conversations helped the company sharpen its long-term strategy: Legora wants to be the platform on which all legal work runs. Unlike previous fundraising rounds, when venture firms were selling themselves to the company, Legora is now pitching investors with access to public markets.

Legora’s CEO said legal AI adoption is accelerating, but more cautiously than adoption of software-development tools. Engineers can experiment and test code, he said, while legal errors may be difficult to identify until they have serious consequences, and assessing work can require a senior practitioner. As a result, he said, legal teams are slower to hand work over entirely to AI agents. He also said a public case in which ChatGPT produced a fabricated legal citation that reached the Financial Times reflects poorly on the field.

Legora’s CEO said its software lets small law firms and in-house legal teams take on more work and more complex matters. He cited large litigation cases, which traditionally required armies of associates to review documents. He also said that in the next quarter, more than half of the company’s incoming business would come from enterprise customers, including legal departments at Salesforce and BMW. In-house teams, he said, are expanding use to contract review and adjacent areas such as compliance and tax. In his view, the tools give smaller practitioners access to capabilities that might otherwise be behind high paywalls.

Legora’s CEO said external legal benchmarks are not very useful for evaluating practical tasks. The company has built an internal library of more than 5,000 real-world tasks with law firms and practitioners, he said. He described model performance as a “jagged frontier”: different models excel at different tasks, so Legora uses multiple models and needs to swap them quickly. He called model routing a core part of the company’s IP. He also claimed that this approach makes Legora significantly more efficient per token than frontier labs and competitors while delivering the same or better performance.

The host reported an OpenAI–Google distribution agreement for OpenAI models through Google Cloud. He said frontier models, including GPT-6 Astra and Codex, would arrive in the coming weeks, and enterprise customers would be able to buy them through Google Cloud Marketplace using existing cloud-spending commitments. The host linked the deal to OpenAI’s expansion to AWS after its exclusive Microsoft distribution arrangement ended.

The host reported that OpenAI acquired OrchestraBio, a platform that helps biopharma teams coordinate the complex process of drug development. No deal terms or integration plans were given on air.

The host summarized comments by Ken Griffin: Griffin said the U.S. is almost certainly in an AI bubble, but also discussed AI-driven disruption and described the period as a potential golden age of entrepreneurship. The host said the remarks sounded less bearish in the context of Griffin’s other comments.

The host said Wop raised $200 million in a Series D round from Sequoia and Shopify. He characterized Shopify’s participation as strategic and compared the company to Shopify for a new category of internet work. The host also praised the team’s ability to adapt to customers, saying its offerings, including clipping and courses, continue to evolve.