Project Glasswing Launches — Anthropic Deploys Mythos to Find Thousands of Zero-Days
Anthropic officially launched Project Glasswing, a major cybersecurity initiative that uses Claude Mythos Preview to identify and fix vulnerabilities in critical software. In testing, Mythos autonomously discovered thousands of previously unknown zero-day vulnerabilities across every major operating system and web browser. In one case, it chained together four vulnerabilities into a working browser exploit and independently achieved local privilege escalation on Linux.
The initiative brings together AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic is committing $100M in usage credits and $4M in direct donations to open-source security organizations. Mythos Preview will not be made generally available — access is restricted to approved partners for defensive security work only.
Google/Broadcom Compute Deal Revealed as 3.5 Gigawatts of TPU Capacity
Details of Anthropic’s expanded infrastructure partnership with Google and Broadcom have emerged via a Broadcom SEC filing. The deal covers 3.5 gigawatts of next-generation TPU capacity — one of the largest AI compute commitments ever made — expected to come online starting in 2027. The vast majority of the new compute will be sited in the United States.
Anthropic trains and runs Claude on a mix of AWS Trainium, Google TPUs, and NVIDIA GPUs, letting it match workloads to the best-suited hardware. The deal extends a November 2025 commitment to invest $50 billion in US computing infrastructure and positions Anthropic for the massive compute requirements of next-generation models like Mythos.
Anthropic Revenue Run Rate Passes $30 Billion
Bloomberg reports Anthropic’s annualized revenue has surpassed $30 billion — roughly triple the ~$9 billion run rate at the end of 2025. The company now counts more than 1,000 business customers spending over $1 million annually, a figure that has more than doubled since February. The growth is driven by enterprise adoption of Claude for agentic workloads, API consumption, and the expanding Claude Code user base.
Claude Code Ships Session ID Headers, /powerup Lessons, and Performance Fixes
The latest Claude Code updates add a new X-Claude-Code-Session-Id header to API requests, letting proxies and monitoring tools aggregate requests by session. Interactive /powerup lessons are now available to help new users learn Claude Code features. The release also includes stronger resume and performance improvements, plus fixes for dialogs, hooks, editing, scrolling, and PowerShell permissions.
Other changes: .jj and .sl directories are now excluded from VCS scanning, privacy controls have been tightened, and session handling is faster across the board.
Simon Willison: Restricting Mythos to Security Researchers “Sounds Necessary”
Developer and blogger Simon Willison published an analysis of Project Glasswing, noting that the decision to restrict Claude Mythos Preview to vetted security researchers rather than releasing it broadly represents a significant shift in how frontier labs think about model deployment. He argued the approach sets a precedent for capability-gated releases where the risk profile of a model, not just its commercial potential, determines who gets access.
Claude Down for Third Consecutive Day — Users Report Login Failures, Chat Errors
Claude experienced renewed connectivity issues on Wednesday, April 8, with hundreds of users reporting login failures, chat errors, and degraded performance across Claude.ai, the mobile app, and Claude Code. This marks the third consecutive day of service disruptions following incidents on April 6 and April 7. Downdetector reports spiked early morning, though Anthropic’s status page showed the issue was identified by 07:59 UTC and a fix was implemented by 09:01 UTC.
The back-to-back outages come at a particularly bad time, coinciding with post-OpenClaw migration traffic and surging demand from agentic workloads. NBC News reported that Anthropic is struggling to meet demand for Claude, connecting the reliability issues to the rapid growth in usage that has pushed the company’s infrastructure to its limits.
NBCNews: Anthropic Struggling to Meet Surging Demand for Claude
NBC News published a deep look at Anthropic’s capacity challenges, reporting that the company is struggling to keep up with demand driven by Claude Code adoption, enterprise agentic workloads, and the broader shift toward AI-powered software development. The report connects the recent outage streak, the OpenClaw subscription crackdown, and the new usage limits as symptoms of the same underlying problem: demand is growing faster than infrastructure can scale.
The Glasswing Paradox: Too Dangerous to Release, Too Important to Shelve
Project Glasswing is simultaneously the most exciting and most unsettling thing to come out of any AI lab this year. A model that autonomously chains together browser exploits and achieves privilege escalation on production operating systems is, by any reasonable definition, a cyberweapon. Anthropic’s response — restrict access, partner with the targets, fund open-source fixes — is the right call, but it also reveals the core tension of frontier AI development.
Meanwhile, $30B in run-rate revenue and three straight days of outages tell the same story from different angles. Anthropic is winning the enterprise race, but the infrastructure is buckling under the weight of its own success. The 3.5GW compute deal with Google and Broadcom is the long-term answer, but it doesn’t come online until 2027. Between now and then, expect more outages, more usage limits, and more frustration from users who are increasingly dependent on a service that can’t always keep the lights on.