What's new in DevEx - April 9

Mythos scored 93.9% on SWE-Bench Verified and autonomously chained four zero-day browser vulnerabilities without human guidance. Rather than release it publicly, Anthropic launched Project Glasswing, giving ~50 organizations including AWS, Cisco, and CrowdStrike gated access for defensive security work only.

Built from the ground up as a natively multimodal reasoning model, Muse Spark uses "thought compression" during training, penalizing the model for excessive thinking tokens to solve complex problems efficiently. In "contemplating" mode it spawns parallel subagents to tackle different parts of a task simultaneously, putting it in direct competition with Gemini Deep Think and GPT Pro's reasoning modes.

The three labs announced they are now sharing live threat intelligence through the Frontier Model Forum to detect and block adversarial distillation attempts. DeepSeek, Moonshot AI, and MiniMax were named directly. Anthropic claims the three collectively generated 16 million exchanges with Claude via roughly 24,000 fraudulent accounts.

Following Claude Code source leak, Claw Code emerged as a cleaned-up open-source reproduction of the same agent harness, built in Python and Rust. It surpassed 100,000 GitHub stars within days despite Anthropic filing DMCA takedowns against every associated repository.

Cursor's biggest interface overhaul replaces single-chat with an Agents Window that runs multiple agents simultaneously across local, cloud, or SSH environments. Design Mode lets developers annotate live browser UI to give agents precise frontend instructions, while /worktree isolates each agent's changes in a separate git branch. Agent Tabs support side-by-side views across multiple repos at once.

Y Combinator-backed Sazabi launched on the thesis that AI agents can extract all operational insight from logs alone, making the metrics and traces layers of traditional stacks redundant. The platform integrates with existing tooling rather than replacing it, targeting early- and growth-stage teams looking to cut monitoring complexity and storage costs. Backers include operators from Vercel, Replit, LangChain, and Browserbase.

🗽 New York

Want to share your dev tool, event, or hot takes?

Submit your story here - we review every submission and highlight the best in future issues!

Till next time,