You know an AI model is serious when it gets two superpowers to pick up the phone.

Anthropic’s Claude Mythos — the model so good at finding software vulnerabilities that it treats global infrastructure like tissue paper — just did something no diplomat, treaty, or summit could manage. It’s pushing the US and China toward actual AI safety talks. Not the performative kind. The “we’re both terrified” kind.

The Model That Broke the Silence

Quick recap if you’ve been living under a rock: Mythos can autonomously discover and exploit software vulnerabilities at a scale that makes seasoned security researchers nervous. The UK’s AI Safety Institute confirmed it can “execute multi-stage attacks on vulnerable networks and discover and exploit vulnerabilities autonomously.” Anthropic says it’s found holes in every major browser and operating system.

This isn’t academic. It’s not a benchmark score. It’s real, exploitable, unpatched flaws in the software running banks, hospitals, and power grids.

Anthropic kept it locked down under “Project Glasswing” — limited access to vetted partners like Apple, Amazon, JPMorgan Chase, and Palo Alto Networks. No public release. No API playground. The grown-up version of “move fast and break things” is “move carefully because things are already broken.”

Two Superpowers, One Problem

Here’s where it gets geopolitically interesting. According to the Los Angeles Times, quiet discussions are underway ahead of Trump’s state visit to China this week to establish an AI “red phone” — an emergency communication channel for AI incidents.

This is a massive reversal. The Trump administration had been aggressively uninterested in AI safety diplomacy. Jake Sullivan, Biden’s former national security advisor who helped set up the first US-China AI dialogue in 2024, says he urged the incoming team to continue those talks. They weren’t interested.

“That’s all changed in the past few weeks,” Sullivan told the Times.

What changed? A model that can find vulnerabilities faster than humans can patch them. And Anthropic CEO Dario Amodei’s warning that China’s frontier AI systems may be only six to twelve months behind Mythos. The window for establishing safety norms is closing fast.

The Strategic Disconnect

The US and China aren’t even running the same race. American AI labs — OpenAI, Google, Anthropic, Meta — are locked in an AGI sprint, betting that whoever gets there first wins everything. China is playing a different game entirely: deep integration of AI into manufacturing, robotics, IoT, and industrial applications.

“The Chinese believe there is no single race, but multiple races,” said Scott Kennedy of CSIS.

This gap has made talks difficult. The Chinese side has historically viewed American hand-wringing about AI existential risk as “a bit abstract.” When Americans showed up to the 2024 Geneva dialogue worried about alignment and loss of control, China wanted to talk export controls.

Mythos might be the first thing making both sides nervous about the same problem at the same time. That’s rare — and potentially productive.

“The Threat Was Already Here”

Not everyone’s buying the panic. Cybersecurity firm watchTowr says researchers are already reproducing Mythos’s findings using publicly available models through clever orchestration. Klaudia Kloc, CEO of Vidoc, claims existing models have been capable of detecting zero-day vulnerabilities at scale for months.

The uncomfortable question: Is Mythos a genuine leap, or has Anthropic — valued at roughly $800 billion and eyeing an IPO — found a spectacular way to demonstrate a capability that was already developing across the industry?

Probably both. Mythos may be the best at what it does, but it didn’t invent the category. What it did is make the threat impossible to ignore. Sometimes the breakthrough isn’t the technology — it’s the demonstration.

Meanwhile, in a California Courtroom

While Anthropic is reshaping geopolitics, OpenAI is having a rougher week. The Musk v. OpenAI trial entered its third week with former chief scientist Ilya Sutskever testifying that he spent about a year compiling evidence of Sam Altman’s “consistent pattern of lying.” Former CTO Mira Murati testified about Altman “saying one thing to one person and completely the opposite to another.”

OpenAI did fire back on the Mythos front — announcing GPT-5.5-Cyber, a cybersecurity-specific model for vetted security teams. The cybersecurity AI arms race between these two companies is accelerating right as both eye IPOs. Nothing like existential software vulnerabilities to juice a valuation.

What Actually Matters

Strip away the corporate drama and geopolitical chess. Here’s what’s left:

AI can now find software vulnerabilities faster than we can fix them. That’s not a future scenario. That’s today. Every company, hospital, bank, and government agency running on digital infrastructure — which is all of them — just got a new threat model.

The “too capable to release” problem is real. Anthropic’s restricted access approach works for now. But as competitors catch up and orchestration techniques make similar capabilities available through public models, that strategy has an expiration date.

Watch the US-China talks. If a real AI safety framework emerges from Trump’s Beijing visit, it would be the most significant international AI governance development in years. If talks collapse, the arms race goes into overdrive.

The AI safety debate stopped being theoretical the moment a model started finding real vulnerabilities in real systems. The only question left: Can our institutions move as fast as the technology that’s forcing their hand?

History suggests no. But Mythos might be scary enough to change the math.