Anthropic's locked-down Mythos leaks
Anthropic's restricted Mythos AI model was accessed by a Discord group shortly after launch, exploiting leaked credentials and URL patterns from the Mercor data breach. The newsletter also covers SpaceX's $60B partnership with Cursor, OpenAI's new Workspace Agents for enterprise teams, and various other AI industry developments.
Summary
The lead story covers the unauthorized access to Anthropic's Mythos model, a cybersecurity-focused AI released on April 10th under 'Project Glasswing' and deemed too powerful for public release. Within days of launch, a private Discord group reportedly accessed Mythos by guessing Anthropic's deployment URL and naming conventions, aided by data leaked in the recent Mercor breach. One group member had vendor credentials through contract work. Bloomberg reported the group has been using Mythos regularly and claims access to other unreleased models, though they deny using it for malicious purposes. The incident is notable because the first unauthorized use of a model that prompted emergency White House meetings came from a casual Discord group rather than a state-level adversary.
The second major story involves SpaceX announcing a partnership with AI coding startup Cursor, including an option to acquire the company for $60 billion. Cursor is guaranteed $10 billion regardless of whether the acquisition is completed, with the full deal contingent on Musk exercising the option before year-end. This move is framed as Elon Musk's attempt to close the gap with Anthropic and OpenAI in the coding AI race, after xAI's Grok has struggled to compete with Claude Code and Codex. The deal provides Cursor access to SpaceX's Colossus supercomputer, addressing compute limitations that had been bottlenecking Cursor's model development.
OpenAI introduced Workspace Agents in ChatGPT, described as an evolution of the 2023 GPT Store concept. These Codex-powered agents are designed for enterprise team workflows, capable of retaining memory, integrating with connected apps, running in Slack, and operating on schedules without user presence. Internal use cases at OpenAI include account research, follow-up drafts, journal entries, and reconciliations.
The newsletter also includes a tutorial on a two-step voice dictation workflow using Typeless and AI coding agents to improve writing quality and build personalized editorial rules over time. Additional brief items cover Google's 8th-generation TPUs, Ideogram's Custom Models for brand-consistent image generation, Google revealing 75% of its code is now AI-generated, Alibaba open-sourcing Qwen3.6-27B, and Odyssey releasing a new world model called Odyssey-2 Max.
Key Insights
- The first known unauthorized use of Anthropic's Mythos — a model serious enough to prompt White House emergency meetings — was carried out not by a nation-state actor but by a hobbyist Discord group using leaked contractor credentials and guessed URL patterns.
- SpaceX's acquisition option for Cursor is structured so that Cursor receives $10 billion guaranteed regardless of outcome, suggesting Musk prioritized securing the asset and its compute relationship over a clean acquisition timeline.
- Elon Musk publicly criticized Cursor by saying it 'was not built right the first time around,' even while xAI was poaching its key engineering leads — and then proceeded to pursue a $60B deal for the same company.
- OpenAI's Workspace Agents are explicitly positioned as a correction to the failed GPT Store, with the company betting that enterprise-grade agentic features and team-sharing infrastructure will succeed where the consumer-facing GPT Store did not.
- Google disclosed that 75% of its internal code is now AI-generated, and separately announced a new generation of TPUs that for the first time separates training and inference into distinct chips, signaling a major architectural shift in how it scales AI workloads.
Topics
Full transcript available for MurmurCast members
Sign Up to Access