Daily Brief

Figure AI livestreamed a team of humanoid robots completing an 8-hour autonomous shift at human walking speeds, with no remote intervention. This is not a lab demo — it's a continuous deployment test run on a public stream where failure would have been visible in real time.

A new Mythos model checkpoint achieves a 60% success rate on a 32-step corporate network penetration task estimated to take a human expert roughly 20 hours. The model completes the full attack chain autonomously, not just individual steps.

Anthropic has discontinued programmatic/API-style use of Claude through subscription products, pushing all automated usage to the official API or enterprise contracts. Simultaneously, Anthropic introduced dedicated monthly Agent SDK credits for paid plans starting June 15, formally separating interactive chat limits from programmatic usage.

Anthropic released Claude for Legal, a vertical product with practice-area-specific plugins and MCP connectors to nine major legal platforms including DocuSign, LexisNexis, and Thomson Reuters. This follows the Claude for Small Business announcement in the same 48-hour window.

OpenAI published a detailed post-mortem on a TanStack npm supply chain attack that affected its tooling, outlining security hardening steps and requiring all macOS Codex app users to update by June 12, 2026 or lose access. The attack vector was a compromised npm package in the TanStack ecosystem.

Xiaomi released MiMo-V2.5-Pro under MIT license: 1.02 trillion total parameters, 42 billion active via MoE routing, with a 1 million token context window. API pricing sits at $70 per 387 million tokens, making the self-hosting economics genuinely competitive for high-volume users.