Karpathy joins Anthropic to lead Claude pretraining; Managed Agents add self-hosted sandboxes

Karpathy's move is the headline talent transfer of the year. After co-founding OpenAI, leading Tesla Autopilot, and a celebrated independent run as an educator-evangelist, he's now reporting into Claude's pretraining org. The signal is unmistakable: Anthropic believes the next jump in Claude quality will come from base-model training improvements, not from stacking more orchestration. Engineers on X are reading it as a vote against the 'just orchestrate' camp that has dominated 2025-2026 product narratives.

The enterprise news is technically meatier. Claude Managed Agents now support self-hosted sandboxes and MCP tunnels, meaning a bank or hospital can keep tool execution entirely inside its own VPC while still letting Anthropic's hosted Claude orchestrate. This closes the biggest gap that had been driving regulated customers toward Bedrock or Azure-hosted models, and it dovetails with KPMG's reported rollout of Claude to 270,000+ employees — a shift from pilot to institutional adoption.

Mythos, Anthropic's cybersecurity-focused preview model, gets a related update: NDAs are being loosened so partners can share vulnerability findings with others. That comes as Mistral courts European banks with an alternative cyber-AI for institutions outside Mythos's tight access list, and as Anthropic confirms a Colossus2/GB200 capacity expansion (HN: 287 pts) — part of the same compute build-out that has Anthropic paying xAI $1.25B/month for GPUs.

Watch next: whether Karpathy's first visible artifact is a Claude 4.5/5 pretraining recipe paper, and whether self-hosted sandboxes show up at AWS re:Invent as a joint Bedrock feature.