Anthropic's Mythos found thousands of zero-days. Intel joined Musk's $25B Terafab. Mercor hit by supply chain attack. OpenAI proposed robot taxes. Perplexity hit $450M ARR on agents. ServiceNow went all-in. Meta locked in $21B for NVIDIA Rubin capacity. Eight stories from the week AI got dangerous — and operational.
Anthropic built an AI model too dangerous to release. Intel joined Musk's $25B Terafab. A training-data giant got supply-chain attacked. Robot taxes proposed. ServiceNow went all-in on agents. Perplexity's pivot drove 50% ARR growth in a month. The EU quietly rolled back AI protections. Meta locked in $21B for NVIDIA Rubin capacity. Eight stories — plus funding rounds, radar signals, and a fake disease that fooled every major LLM.
Anthropic Built an AI That Finds Zero-Days in Every Major OS.
Claude Mythos Preview, Anthropic's unreleased frontier model, has identified
of previously unknown zero-day vulnerabilities in every major operating system, every major web browser, and a range of critical software — including a
flaw in FFmpeg that survived five million automated test runs. The model wrote a browser exploit chaining four vulnerabilities, escaping both renderer and OS sandboxes. It converts
of discovered vulnerabilities into working exploits in Firefox's JavaScript shell. Anthropic won't make Mythos generally available. Instead, it launched Project Glasswing — a restricted partnership with Amazon, Apple, Google, Microsoft, NVIDIA, and Palo Alto Networks to secure critical infrastructure before models with similar capabilities proliferate.
First credible demo that AI capability has crossed a security threshold. Anthropic made the right call restricting access, but the window is narrow — other labs will reach this within months.
Intel Joins Musk's $25B Terafab Megaproject.
Intel signed on as foundry partner for Terafab, the joint venture between Tesla and SpaceX. The partnership claims it will deliver one terawatt of annual AI compute from an Austin megafab. Intel is contributing its 18A process node, a
technology that's the most advanced U.S.-made semiconductor process. Total project cost:
. Analysts remain skeptical the budget will match the ambition, but the deal represents the biggest external customer win for Intel CEO Lip-Bu Tan. It's a direct challenge to NVIDIA's compute dominance and a bet-the-company moment for Intel's foundry business.
One-terawatt target is aspirational. But Intel's 18A is real, and Tan's desperation to prove the foundry model works makes this credible. If it delivers even 20% of capacity, it reshapes the U.S.-China compute balance.
AI Training Data Goldmine Breached Via Open-Source Dependency.
startup that provides AI training data to the world's top frontier labs, confirmed a devastating supply chain attack through LiteLLM, an open-source library downloaded millions of times daily. The hacking group TeamPCP planted malicious code by compromising the maintainer's credentials. One major customer immediately froze its partnership.
of data were reportedly stolen, including database records, source code, and customer records. The Lapsus$ group claimed involvement. This is the attack the industry was warned about — AI companies built their competitive moats on training data, and a single compromised open-source dependency just exposed the entire supply chain.
This is the attack the industry was warned about. AI companies built their moats on training data — and a compromised open-source dependency just exposed the entire supply chain. Every company using LiteLLM should be running forensics right now.
OpenAI Proposes Robot Taxes, Wealth Funds, Four-Day Week.
titled "Industrial Policy for the Intelligence Age" with