Welcome to the agent platform research briefing for Saturday, June 6th, 2026.

**Anthropic calls for coordinated global pause on AI development** — On June 5th, Anthropic published a warning that AI systems may soon be capable of recursive self-improvement without meaningful human involvement, and called on major AI labs to consider a coordinated, verifiable pause in development if risks rise. The company says it will convene policymakers, researchers, civil society, and other AI companies to discuss safeguards. The announcement came alongside reports that something "unsettling" has been happening with Claude — Dario Amodei himself said humans risk losing control over AI systems in the very near future if a concerning trend continues. It's an extraordinary move: the company racing to IPO at a $965 billion valuation is simultaneously arguing that someone should pull the emergency brake.

**Anthropic engineers embedded inside NSA for offensive cyber ops, while company simultaneously sues the Pentagon** — In a remarkable contradiction reported by the Financial Times and TechCrunch on June 5th, approximately six Anthropic engineers are embedded at the NSA as forward-deployed staff, adapting the Claude Mythos model for specific operational cyber applications. The NSA is reportedly using Mythos for offensive cyber operations. Yet Anthropic is simultaneously suing the Defense Department over Secretary Hegseth's June 3rd decision to uphold the supply-chain risk designation barring the company from DoD contracts. Anthropic has refused to allow Claude to be used for mass domestic surveillance or fully autonomous lethal weapons. The legal fight now heads to a three-judge panel at the D.C. Circuit Court of Appeals.

**Project Glasswing expands to 150+ organizations plus ENISA, Claude Security goes broader** — On June 2nd, Anthropic expanded its Project Glasswing cybersecurity initiative to approximately 150 additional organizations across more than 15 countries, adding power utilities and water treatment facilities. On June 1st, the EU's cybersecurity agency ENISA became the first EU institution admitted to the program. Claude Security is now handling codebase scans and patch suggestions for the expanded participant base — the controlled-access Mythos model doing active vulnerability hunting for entities that guard critical infrastructure.

**Claude Code Auto Mode lands on Pro plan, ZoomInfo and Infoblox bring MCP to enterprise data** — Three smaller but structurally significant developments: Claude Code's Auto Mode with background safety checks is now available on the $20 Pro plan with Sonnet 4.6 support, bringing autonomous coding to the mass developer tier. ZoomInfo released a native MCP connector giving Claude access to 500 million verified contacts through its GTM.AI layer. And Infoblox launched Infoblox IQ — the first production system built on DNS-AID, a new Linux Foundation standard for AI agent discovery using DNS, plus an MCP server specification. Together these show MCP moving past developer tools and into real enterprise data infrastructure.

**GPT-5.6 tracker — Polymarket odds shift to mid-June** — Benchmark leaks attributed to internal OpenAI testers began circulating this week, with credible leaker reports referencing checkpoints called Iris-alpha. Polymarket consensus now points to a mid-June GPT-5.6 rollout, though OpenAI has made no official confirmation. Expected improvements include enhanced reasoning for multi-step agentic workflows and longer context windows. If it lands between the Anthropic IPO filing and SpaceX's June 12th IPO debut, the June AI financial calendar gets even more crowded.

That's the briefing for Saturday, June 6th.