Cybersecurity Panic, Nationalizing AI, The Model Squeeze & Can't Own Anything

A PHP vulnerability reportedly known since 2004 had no previously known exploit until a ChatGPT model found a way to exploit it months before public disclosure. OpenAI’s Patch the World, a security researcher, and Wiz independently found the flaw and coordinated disclosure and patching. The exploit can give an attacker a command shell on a vulnerable server, potentially opening access to databases and stolen data. Speakers estimated that roughly 70% of websites by site count use PHP, including sites running WordPress and Drupal.

Cloudflare CEO Matthew Prince reportedly saw hundreds of exploit attempts per hour before disclosure, tens of thousands in the following 45 minutes, and millions per hour later. Greg Brockman estimated researchers did about 80% of the vulnerability-finding work, with more advanced models finding additional flaws after researchers identified where to look. Railway said every secret deployed on its platform might be compromised, and its service and hosted sites went down, but speakers did not know whether the PHP flaw caused the outage or whether Railway used PHP. They estimated about $1 million in crypto was taken using Railway credentials.

The incidents prompted predictions that AI will make it easier to scan the internet for vulnerable sites and test them for valuable targets, with older companies and infrastructure especially exposed. Speakers expected some websites to go down and survivors to adopt stronger, standardized defenses, but said massive society-wide failures were unlikely within six to twelve months. They also argued that ordinary patching and breach detection do not require superhuman models: even models such as GLM 5.3 can patch many vulnerabilities. Rather than nationalizing OpenAI on cybersecurity grounds, governments could fund GPU purchases at market prices for public-infrastructure security work often left to unpaid volunteers.

Nationalization could still be considered for other reasons, including national security or AI’s importance to warfare and civilizational competition. Speakers disagreed about whether a serious data-center downturn would trigger government ownership: a financial crisis might prompt intervention, but a smaller wipeout or private-credit losses might not. They cited the U.S. government’s 61% stake in General Motors after TARP, which it sold over the following five years, as an example of intervention that did not become permanent ownership. One speaker expected people to move on after a 70% valuation decline; another thought a bailout would follow. The case for rescuing OpenAI and Anthropic would be weaker if Google and xAI were similarly advanced.

The economics of building models remain strained. OpenAI and Anthropic were estimated to account for about 30% of data-center capital expenditure, while some seed-stage AI labs valued at $3–4 billion were described as having little beyond former OpenAI interns and a rough white paper. Step Fun and Kimi were characterized as capable but deeply cash-flow negative, sometimes selling Chinese models below inference cost. DeepSeek reportedly has positive inference margins but may still lose money per model generation after training costs. Speakers disagreed over whether Chinese frontier labs are state-owned; they noted that the Chinese Communist Party pays attention to them, while one speaker distinguished them from Alibaba and Tencent and pointed to mainland listing requirements for profitability.

Anthropic was described as profitable, particularly through API margins, and as dominating AI spending even on OpenRouter; speakers cautioned that token share alone is a poor measure of market position. Providers may succeed at enormous scale on the low end or by offering top-tier intelligence, leaving the middle exposed. OpenAI lets customers spend credits on models including Kimi K3 and GLM 5.3, raising the prospect of an enterprise control plane that routes spending across models. But prioritizing that role over frontier leadership could widen its gap with model builders, while its consumer-chat focus has already made it harder to catch Anthropic in enterprise and coding. Meta’s stock fell about 9% after it raised the low end of its capital-expenditure forecast without giving a 2027 figure; a speaker also questioned how it would monetize the $100 billion it was said to have spent building a model.

Questions about who owns AI’s gains surfaced in Anthropic’s plan for a public benefit fund using some of its shares and Elon Musk’s statement that he would give one SpaceX share to every U.S. child; speakers did not equate these proposals with nationalization. They also debated a reported Visa–Fireblocks acquisition, which both companies left unconfirmed. Fireblocks’ last round reportedly valued it at $8 billion, while hypothetical sale values of $1.7 billion to $2.5 billion could fall below an estimated $750 million to $1 billion preferred stack and wipe out common holders. Haseeb said he could not assess a deal without financials and that Visa stock or an earn-out could change the headline price; he would probably accept Visa stock given the weak IPO market and limited liquidity alternatives. Fireblocks was described as a custody business with roughly $150 million to $250 million in ARR estimates, not a cybersecurity company.