Anthropic Mythos Creates Exploits in 31 Minutes

Frontier AI has crossed from finding vulnerabilities to weaponizing them, shrinking the patch gap from weeks to hours.

June 9, 2026 | Reading time: 8 minutes | Issue #184

Anthropic's frontier red-team model, Mythos Preview, can turn publicly disclosed software vulnerabilities into working exploits in hours instead of weeks, according to research the company published on June 8 and shared first with Axios. In one test, Mythos generated a proof-of-concept exploit for a Windows kernel vulnerability within 31 minutes. Across 21 kernel bugs, the model caused a system crash in 18 and produced 8 distinct exploits. The longest took 5.7 hours. On Firefox, Mythos built 8 working code-execution exploits across 18 security patches.

The research establishes a new baseline: N-day vulnerabilities — flaws already known to defenders but not yet patched — can be weaponized by AI faster than most enterprises can distribute fixes. Anthropic estimates each privilege-escalation exploit cost roughly $2,000 in API credits. At that price, the attacker surface expands from sophisticated nation-states to any funded threat group with API access.

Cybersecurity has always operated on asymmetry: attackers need one opening, defenders must cover every door. Mythos narrows the window further. Patches now require not just distribution but pre-emptive hardening, because the time between disclosure and weaponization is collapsing. Anthropic says the research is meant to improve patching prioritization. The problem is that the same capability is available to attackers. Open-source models are already finding bugs at Mythos's level, and the Trump administration's new AI security executive order, signed June 2, does not yet mandate controls on models that weaponize rather than merely discover vulnerabilities.

The real question is whether cybersecurity vendors can build automated patching pipelines fast enough to match a generation of AI that writes exploits faster than humans can review them.

Source: Axios

Apple Overhauls Siri With Google Silicon and NVIDIA GPUs

Apple WWDC 2026 unveiled Siri AI, a rebuilt assistant running on Apple's third-generation Foundation Models trained with Google Gemini outputs and running on NVIDIA GPUs inside Google Cloud data centers. Craig Federighi confirmed the models are Apple's own, not Google's consumer-facing Gemini, but the training data and recipes came from Google. The most powerful on-device model requires an iPhone 17 Pro or newer with 12GB RAM. On the server side, Apple is expanding Private Cloud Compute to third-party infrastructure for the first time.

The move marks a departure from Apple's chip-silos strategy — the company is acknowledging that its own silicon cannot yet handle frontier inference workloads alone. Siri AI will not launch in the EU or China; Apple cited the Digital Markets Act and regulatory uncertainty. The iOS 27 developer beta is available now, but consumer features require waitlist approval.

Sources: Apple Newsroom, CNBC

Cursor Hits $4 Billion Annualized Revenue Ahead of SpaceX Deal

Cursor, the AI coding assistant, crossed $4 billion in annualized revenue in the first week of June, up from $3 billion in April and $2 billion in February, Forbes reported on June 8. The startup is reportedly preparing for an acquisition by SpaceX. The jump from $2 billion to $4 billion in four months makes Cursor one of the fastest-growing software companies on record. The SpaceX angle suggests Elon Musk is consolidating AI tooling ahead of xAI's broader platform push.

Source: Forbes

Trump Administration Moves to Block State AI Laws

The White House is relaunching efforts to preempt state-level AI regulation, Axios reported on June 8. Senator Marsha Blackburn is leading negotiations and pushing to include the Kids Online Safety Act within an AI preemption package. The move would override laws in states like California and Colorado that have already enacted AI safety and transparency rules. The administration's argument is that a patchwork of state laws creates compliance costs that advantage large incumbents. The counter-argument is that a federal vacuum has already allowed unchecked model deployment, and preemption would remove the only safety guardrails that currently exist.

Source: Axios

Compute Watch

Google has placed an order with Intel to manufacture more than 3 million TPUs in 2028 using Intel's 18A process, The Information reported on June 8. Separately, NVIDIA is testing Intel's 18A node for a future processor. Intel's stock jumped 11% on the news. For Google, the order is a hedge against foundry concentration: TSMC currently produces the bulk of Google's AI accelerators. For Intel, landing both Google and NVIDIA validation trials is existential — the company needs high-volume AI contracts to justify its foundry pivot. The 2028 delivery timeline means these chips will hit the market as Google is expected to be on its sixth TPU generation, likely competing with whatever replaces NVIDIA's Vera Rubin architecture. The larger point is that Big Tech is actively diversifying its silicon supply chain, not waiting for Congress to resolve chip subsidies.

Source: The Information

Builder's Corner

Apple announced a Foundation Models framework for developers and a Core AI framework alongside Xcode enhancements aimed at agentic coding workflows. Developers will be able to describe Shortcuts with natural language and have Apple Intelligence generate automations. The agentic architecture is an acknowledgment that Apple cannot build every AI feature itself. By opening agentic tools to third-party developers, Apple is positioning iOS 27 as an operating system for AI agents rather than merely an AI-enhanced operating system. The risk is execution: Apple's developer frameworks have historically been powerful but tightly controlled, and agentic AI requires looser constraints than Apple typically tolerates.

Source: MacRumors

India Lens

Indian quick-commerce startup Zepto filed for a $836 million IPO on June 8, Bloomberg reported, at a $7 billion valuation. While Zepto is not an AI company, its filing signals that India's startup pipeline is maturing beyond SaaS clones into consumer infrastructure with real unit economics. On the AI front, a recent Rest of World analysis framed India's "frugal AI" movement — led by Sarvam AI and Krutrim — as a blueprint for resource-strapped nations. The thesis is simple: India cannot outspend the US or China on training runs, so it optimizes for inference efficiency and domain-specific models in regional languages. With TCS already reporting $2.3 billion in annualized AI services revenue, India's play is not frontier model dominance but implementation at scale. The Zepto IPO and the AI services numbers suggest the country is building the middle layer — deployment, integration, and distribution — while others fight over the model layer.

Source: Rest of World

Eastern Front

The US Department of Defense designated Alibaba, BYD, and Baidu as entities supporting the Chinese military, Bloomberg reported on June 8. The designation adds pressure to an already strained trade relationship and follows the Bureau of Industry and Security's May 31 guidance that expanded AI chip licensing requirements to Chinese-headquartered firms operating outside mainland China. The net effect is a tightening noose on offshore workarounds: Chinese AI labs that tried to route compute through Singapore or Dubai now face the same licensing scrutiny as mainland operations. Apple's announcement that Siri AI would skip China at launch, citing regulatory uncertainty, further illustrates that US-China tech bifurcation is accelerating from hardware into consumer software.

Source: Bloomberg

The View

Anthropic's Mythos, Apple's multi-cloud Siri, and the DOD's Alibaba designation describe a single pattern: AI infrastructure is fragmenting by capability and by geography. On one track, frontier models are acquiring destructive skills — exploit generation, autonomous coding — faster than governance structures can absorb them. On another, the physical infrastructure that runs those models is splitting into US-aligned and China-aligned stacks, with Apple now explicitly choosing Google Cloud and NVIDIA over its own in-house silicon and Chinese regulatory approval. The tension is that weaponized AI does not respect borders. A model that writes kernel exploits in 31 minutes can be deployed from anywhere, by anyone with API access. Geographic decoupling of hardware does not decouple the threat. Policymakers are still thinking in terms of export controls on chips; they should be thinking in terms of access controls on capabilities.

The Miss

A study published in the New York Times on June 8 estimates that the iPhone's diffusion explains 33% to 52% of the decline in US births from 2007 to 2011, with the most acute effects among women aged 15 to 24. The finding is not AI-specific, but it frames a deeper question the industry avoids: the most consequential technologies are not the ones that write code or generate exploits, but the ones that reshape human behavior at population scale. AI's second-order social effects — screen time, isolation, attention capture — are being engineered by the same labs that publish security research and sell agents. Coverage of those effects remains sparse.

Source: New York Times

Pull Quotes

"We see Siri not as a separate chatbot, but rather as an integral but conversational tool." — Craig Federighi, Apple, via Tom's Guide

"Patching a system isn't always as easy as downloading a software update." — Anthropic research, via Axios

"Europe has two years before becoming America's AI vassal state." — Arthur Mensch, Mistral CEO, via Business Insider

"I can confirm that Siri AI works in the UK, which doesn't fall under the EU restriction. Source: me. I'm using it." — Connor Jewiss, via X

Reads & Links

• Axios: Anthropic's Mythos can exploit new flaws in hours
• Bloomberg: Pentagon accuses Alibaba, Baidu, BYD of aiding China's military
• The Information: Google, NVIDIA consider Intel as backup chip manufacturer
• Forbes: Cursor passes $4 billion annualized revenue
• Rest of World: India's frugal AI models blueprint
• Bloomberg: Zepto files for planned India IPO

Out

The patch gap was already a losing battle. Now it is measured in minutes.

By Neo