The AI Brief

OpenAI published its Frontier Governance Framework on May 29, stating that it "explains how our safety and security practices align with emerging legal requirements, including California's Transparency in Frontier AI Act and the EU AI Act's Code of Practice for General Purpose AI." The framework is not a new safety methodology, but rather a regulatory translation layer.

OpenAI's coverage spans cyber offense, CBRN (chemical, biological, radiological, nuclear) misuse risks, harmful manipulation, and loss-of-control scenarios. Beyond risk categories, the framework also documents model reporting procedures, security risk management protocols, incident response workflows, and mechanisms for incorporating external expert input. The framework defines systemic risk as foreseeable material risks of severe harm, specifically including scenarios where a model contributes to more than 50 fatalities or causes $1 billion in property damages from a single incident. To maintain accurate risk baselines, OpenAI solicits input from external domain experts and independent third-party evaluators; these external experts help stress-test safeguards for models approaching a new risk tier.

Under the EU AI Act provisions, the company commits to evaluating whether to update its Safety and Security Model Report for its most capable models every six months. Updates are considered required if a model's capabilities materially change through post-training or if integrations into internal systems increase risk. The framework was published alongside a companion "Trusted Access for Cyber" initiative that commits $10 million in API credits to accelerate cyber defence work and establishes an identity-based access tier for security professionals. The document also explicitly defines a systemic risk threshold, a detail that gives enterprise compliance teams a concrete benchmark for assessing their own deployment obligations under California's SB 53, which was signed in September 2025.

Anthropic's annualized revenue run rate hit $47 billion this month, up from about $10 billion a year earlier, according to CNBC. That revenue line is the part that flipped Wall Street: investors had questioned whether any of these labs could turn frontier models into a real business instead of a cash furnace. Anthropic's rise runs through code, not chat. Claude Code, the company's AI coding assistant, is the primary driver of the acceleration.

On May 28, Anthropic said it raised $65 billion in fresh funding at a $965 billion valuation, vaulting past OpenAI to become the most valuable AI startup in the world, according to CNBC. The Series H round was led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. It lifts Anthropic to a $965 billion post-money valuation, ahead of OpenAI's $852 billion mark from late March. One market estimate puts Anthropic at 40% of enterprise LLM revenue, with OpenAI at 27% and Google at 21%. That figure is from Menlo Ventures and carries vendor-interest caveats.

The revenue figure is self-reported by Anthropic via CNBC and has not been independently audited. Both the run-rate figure and the enterprise market-share estimate should be treated as directional rather than precise.

Disclosure: Anthropic, mentioned in this item, is the company that develops Claude, which generates this brief.

Dreaming extends memory by reviewing past sessions to find patterns and help agents self-improve. The platform also gained outcomes, multiagent orchestration, and webhooks for developers building with Managed Agents. Together, these updates make agents more capable at handling complex tasks with minimal steering. Dreaming is a scheduled process that reviews agent sessions and memory stores, extracts patterns, and curates memories so agents improve over time. Operators decide how much control they want: dreaming can update memory automatically, or changes can be reviewed before they land. Dreaming surfaces patterns a single agent cannot see on its own, including recurring mistakes, workflows that agents converge on, and preferences shared across a team. It also restructures memory so it stays high-signal as it evolves.

Teams are already using dreaming in production. Harvey, the legal AI platform, uses Managed Agents to coordinate complex legal work like long-form drafting and document creation. With dreaming, their agents remember what they learned between sessions, including filetype workarounds and tool-specific patterns; completion rates went up roughly 6x in their tests. The outcomes feature, promoted to public beta alongside dreaming, spawns a separate grader agent that evaluates output against an operator-defined rubric without access to the original agent's reasoning, which closes a specific failure mode: agents rationalising flawed work. Anthropic reported up to 10 percentage points of task success improvement in internal testing.

The MCP tunnels feature, added May 19, allows Claude Managed Agents users to route services through a private network: agents reach MCP servers inside a private network without exposing them to the public internet. Internal databases, private APIs, knowledge bases, and ticketing systems become tools agents can call. A lightweight gateway makes a single outbound connection, with no inbound firewall rules, no public endpoints, and traffic encrypted end-to-end. Dreaming is available in research preview; outcomes, multiagent orchestration, and MCP tunnels are in public beta. Note that Claude Managed Agents sessions are not currently eligible for Zero Data Retention or HIPAA BAA coverage, a material constraint for healthcare operators.

Disclosure: Anthropic, mentioned in this item, is the company that develops Claude, which generates this brief.

The Colorado AI Act, currently slated to come into effect on June 30, 2026, will place substantial new responsibilities on AI developers and deployers, including requirements to undertake reasonable care to avoid algorithmic discrimination, to develop a risk management policy and program, to implement notices, and to conduct impact assessments. Proposed amendments include removing the stand-alone affirmative duty to "avoid algorithmic discrimination" that appeared as an explicit independent requirement in the original SB 205, with these proposed changes signalling state action moving away from a broad "high-risk AI" framework toward a narrower, decision-focused model. As of publication, it remains unclear whether those amendments will be enacted before June 30.

President Trump tasked the Attorney General with coordinating litigation efforts against state measures that impose what the administration characterises as innovation-limiting requirements. In January 2026 the Attorney General announced an AI litigation task force. The executive order also allows federal funding and infrastructure support to be conditioned on state alignment with national AI policy. The March 20 White House framework is not a binding document and does not by itself impose new legal obligations or direct agencies to take specific regulatory actions. Instead, it outlines a series of recommended policy approaches for Congress to consider in drafting comprehensive federal AI legislation.

State AI laws remain in effect unless and until Congress enacts new legislation or agencies or other stakeholders succeed in legal challenges to state laws. The EU AI Act presents a parallel deadline: broader obligations become applicable on August 2, 2026. Operators with EU exposure face two compliance clocks running simultaneously.

Grok Build runs up to eight parallel AI agents simultaneously, each working through a three-stage workflow: plan, search, and build. What sets it apart from other tools is Arena Mode, an automated evaluation layer that scores and ranks competing outputs before a developer ever reviews them. The underlying model, grok-code-fast-1, was built from scratch, separate from the Grok 4 lineage, with a training corpus heavy on programming content and post-training focused on real-world pull requests and coding tasks. It scores 70.8% on SWE-Bench Verified and is priced at $0.20 per million input tokens, a notably competitive price point compared to what developers are paying for Claude Code or Codex CLI.

The tool is local-first, meaning no source code is transmitted to xAI's servers. For teams working with proprietary codebases or in regulated industries, that is a meaningful design choice. Released in early access on May 14, Grok Build 0.1 is a purpose-built coding model trained specifically for agentic workflows, designed to take multi-step actions rather than just answer questions. It supports text and image inputs, outputs text, and carries a 256,000-token context window. As of publication, broader availability remains on a waitlist and access is limited to SuperGrok and X Premium Plus subscribers.

The Grok Build launch feels less like a product announcement and more like a strategic statement: xAI needs a credible foothold in enterprise developer workflows, and coding agents are one of the clearest paths to get there. SWE-Bench Verified scores from third parties have not yet been independently replicated for grok-code-fast-1; treat the 70.8% figure as vendor-reported until confirmed.