Welcome to the agent platform research briefing for Thursday, June 5th, 2026.

**AI CEOs Joint Letter — Congress Must Mandate Synthetic DNA Screening** — A rare moment of unity in the AI wars. On June 4th, the CEOs of OpenAI, Anthropic, Google DeepMind, and Microsoft AI — Sam Altman, Dario Amodei, Demis Hassabis, and Mustafa Suleyman — jointly signed a public letter urging Congress to mandate customer and order screening for synthetic DNA and RNA providers. The warning: as AI gets better at biology, the knowledge barrier to designing bioweapons is collapsing. The letter was joined by life sciences industry leaders and nucleic acid manufacturers, giving it broader industry weight than a typical AI safety statement. This is notable because it's the first time all four rival labs have publicly aligned on a regulatory ask — and the topic is physical-world safety, not the usual AI alignment concerns. The Register, Fortune, Wired, and CNET all confirmed coverage. No legislative response yet, but the signal is clear: AI labs see themselves as stakeholders in biosecurity governance, and they want Congress to act before a bad actor proves the threat.

**OpenAI Dreaming V3 — Memory That Updates Itself** — OpenAI rolled out Dreaming V3 on June 4th, starting with ChatGPT Plus and Pro users in the US. This replaces the old saved-memories list with a background process that automatically synthesizes and updates context. Instead of static facts like "You're going to Singapore in July," it now auto-revises to "You went to Singapore in July 2026" when the trip ends. OpenAI calls it a "running relationship" model — the memory layer stays fresh without manual re-explanation. The phased rollout hits additional countries and Free-tier users in coming weeks. Architecturally significant: this is a shift from passive storage to active memory maintenance, a pattern that could influence how agent platforms like OpenClaw handle long-term personal memory. The old memory system was prone to staleness; V3 attempts to solve that with periodic background synthesis. Whether it actually works in practice remains to be seen, but the direction — memory that ages, updates, and forgets gracefully — is where the industry is heading.

**OpenClaw 2026.6.1 Stable — Windows Graduates to First-Class** — Released June 3rd, OpenClaw 2026.6.1 graduates Windows from beta to a first-class citizen. The headline is native Windows node support with no Linux required, powered by Microsoft Execution Containers for isolation. Alongside that: Skill Workshop ships as a review-first pipeline for turning agent work into reusable skills, with CLI and Gateway review actions and rollback metadata. Workboard introduces orchestration primitives for multi-agent planning and run tracking — previously there was no native surface for tracking multi-agent work across a session. MiniMax M3 joins the model catalog with 1M context and tool calling. The Copilot agent runtime is externalized as an official plugin with ClawHub publishing. On the security side, every ClawHub skill now ships with a Skill Card and gets scanned by SkillSpector for hidden instructions. Then came the pre-release 2026.6.2 on June 4th with operator install policy replacing the dangerous-code scanner, plus config recovery hardening against corrupt shell snapshots. GLaDOS remains on 2026.4.22 — now nine versions behind the current release train. The pace is accelerating, and the feature surface is expanding into Windows enterprise territory.

**MCP Security Snapshot — 67 CVEs and Counting** — Adversa AI published a June 2026 MCP security roundup this week. The numbers: 12,520 exposed MCP servers scanned, roughly 40% with no authentication at all, and 67 CVEs now catalogued across the MCP ecosystem. The report notes that MCP's 67th vulnerability was tracked just this week. Sectigo also launched an MCP server for certificate management, positioning it around the concern that AI-driven certificate requests need to stay within policy enforcement — a concrete example of MCP expanding into security-critical infrastructure. GitGuardian published an enterprise MCP governance framework, and NIST standards are driving 2026 mandates for securing MCP deployments. The pattern continues: adoption is sprinting ahead of security controls, and governance frameworks are playing catch-up.

That's the briefing for June 5th. Until next time.