Mitchell Hashimoto just quit GitHub after 18 years. His journal shows outages on almost every single day this past month. Anthropic quietly doubled Claude Code costs. And an AI model called Mythos found zero-day vulnerabilities so dangerous that it can't be released publicly. The AI industry is eating its own infrastructure.
What's Breaking
HashiCorp Co-Founder Quits GitHub: "No Longer a Place for Serious Work" Mitchell Hashimoto kept a journal this past month — an 'X' on every date where a GitHub outage blocked his work. Almost every day has one. On the day he wrote his farewell post, he couldn't do PR reviews for two hours due to a GitHub Actions outage. GitHub CTO Vlad Fedorov published his second availability update in six weeks, admitting AI is driving "record acceleration" the platform can't handle. After 18 years, one of the most respected infrastructure engineers in the world is gone.
Anthropic Quietly Doubles Claude Code Costs: $6→$13/Developer/Day On April 15, Anthropic updated its website to more than double Claude Code cost estimates. Average: $6→$13/developer/day (117% increase). P90: $12→$30 (150% increase). Their head of growth admitted: "Our current plans weren't built for this." At enterprise scale, you're looking at $150-250/developer/month. The culprit? Opus 4.7 costs 2x+ what Sonnet 3.7 did.
Business Insider, Yahoo Finance
Enterprise AI Pipeline Burns Thousands in Tokens for Incomplete Results A developer documented building an end-to-end AI bug-fix pipeline (Jira → diagnosis → fix → review → CI/CD). One test run: 117 tool calls in a single turn before the API aborted. A 12+ node pipeline where any single tool failure breaks the entire flow. Context accumulation causes judgment drift. Sessions are stateless in enterprise environments. The verdict: "From a pure cost standpoint, no — the per-run token cost is still high."
Dev.to — Enterprise AI Pipeline Reality
Top AI News
Mythos: First AI Model Too Dangerous for Public Release Anthropic's Mythos model found zero-day vulnerabilities in critical infrastructure, including "decades-old bugs" that nobody had discovered. It's too dangerous to release publicly. Anthropic launched Project Glasswing — coordinated patching with AWS, Apple, Microsoft, Google, and Cisco. The NSA is already using it. The White House is now drafting guidance to bring Anthropic back into federal procurement after previously restricting it.
Poolside Laguna: Open-Source Coding Model Runs on a Mac Poolside released Laguna M.1 (72.5% SWE-bench) and XS.2 (68.2% SWE-bench, Apache 2.0, runs on a Mac with 36GB RAM). First coding model family trained with explicit agent RL methodology. The XS.2 at 33B/3B active parameters is the strongest open-weight coding model at its size — and it fits on your laptop.
OpenAI Symphony: CI/CD for AI Agents OpenAI open-sourced Symphony, an agent orchestration spec (Apache 2.0) that coordinates coding agents from issue tracker to PR. Tens of thousands of GitHub stars within weeks. If it catches on, it becomes the standard for how AI agents collaborate in software development.
Amazon: All AI Output Must Be Human-Reviewed Amazon Stores director Steve Tarcza: "Every mutating step that an AI might do requires a human to approve it." This directly contradicted AWS's own keynote hyping agentic AI as "magic." The gap between marketing hype and enterprise reality has never been wider.
The infrastructure paradox: AI tools depend on GitHub, but AI-driven load is breaking GitHub. Vendors promise "magic," but Amazon mandates human review of every AI action. The hype curve has peaked — reality is setting in.
Papers That Matter
Poolside Laguna: Agent RL Training for Coding (Poolside AI) First publicly released coding model trained with explicit agent reinforcement learning — the model learns by actually doing coding tasks in a terminal environment, not just from code snippets. Hits 72.5% on SWE-bench and the smaller XS.2 runs locally at 68.2%. This is the blueprint for how coding agents should be trained — in the environment where they'll actually work.
AltTrain: 3-Step Reasoning Prevents Destructive Actions (KAIST) Three-step reasoning structure (understand → assess harm → conditionally reason) reduces harmful outputs from 83.5% to 4.8%. Training takes 60 minutes on a single GPU. With Mythos finding zero-days and GitHub buckling under AI load, this kind of built-in safety reasoning is becoming non-negotiable.
What This Means For You
Mitchell Hashimoto quitting GitHub isn't a personal decision — it's a canary in the coal mine. AI-driven load is breaking the infrastructure that AI tools depend on. GitHub can't handle it. Claude can't stay up. OpenAI had its own GPT-4o-mini outage. When your AI coding agent depends on GitHub for PRs, and GitHub is down because of AI load, you've got a circular dependency that breaks your entire workflow. Build redundancy now — mirror critical repos, have fallback CI/CD, and don't bet your development pipeline on a single platform.
Anthropic doubling Claude Code costs while their reliability tanks is a brutal combination. You're paying more for less. The enterprise math is stark: $150-250/developer/month for a tool that might be down when you need it most. Poolside's Laguna XS.2 running locally at 68.2% SWE-bench with Apache 2.0 licensing isn't just a cool demo — it's an escape hatch from vendor dependency.
The Mythos situation changes the conversation about AI safety from theoretical to urgent. This model found real zero-days in critical infrastructure. Anthropic did the right thing by coordinating disclosure through Project Glasswing. But if Anthropic has this capability, others will too — including actors who won't be as responsible. The time to invest in AI safety infrastructure isn't after the next incident. It's now.
Written by The AI Architect team at Atobotz