The AI Brief

On June 11, 2026, Google DeepMind announced a research funding call of up to $10M, jointly backed by Schmidt Sciences, the Cooperative AI Foundation, ARIA (the UK government's Advanced Research and Invention Agency), and Google.org. The call targets four research areas: sandboxes and testbeds for realistic multi-agent environments, the science of agent networks, infrastructure for safer agent-to-agent communication, and oversight and control mechanisms. Applications close August 8, 2026, with awardees announced in autumn.

Rohin Shah, who directs AGI safety and alignment research at Google DeepMind, framed the problem as a structural shift: soon, millions of AI agents built by different organizations will interact across digital environments, communicating, negotiating, and transacting with one another. Shah told MIT Technology Review that the mass-market arrival of agents that can carry out tasks without human oversight and follow instructions from other agents "creates a whole new class of risk." The concern is not hypothetical: agent-based tools were a centerpiece of Google I/O in May 2026, and Microsoft's multi-agent framework is already in production-grade deployment.

Critics within the research community acknowledge the importance of the initiative while noting its scope limitations. One researcher quoted by MIT Technology Review warned that safety researchers can be drawn toward exotic hypothetical risks while overlooking "boring problems that are already here." The fund does not specify what share of the $10M will go to empirical versus theoretical work, and no lab has committed to incorporating the findings into its deployment decisions.

Google Research announced Gemini-SQL2 on June 12, 2026. Built on Gemini 3.1 Pro, it posted 80.04% execution accuracy on the BIRD single-model leaderboard. BIRD covers 12,751 question-SQL pairs across 95 databases in 37 professional domains and scores SQL that actually executes and returns the correct result, not SQL that merely looks syntactically correct. Human performance on BIRD stands at 92.96%, leaving a 12.92-point gap above Gemini-SQL2. The prior leaderboard top was Google's own Gemini-SQL at approximately 77.2%, with AWS's Q-SQL following at roughly 76.5% and Anthropic's Claude Opus 4.6 at approximately 70.1% on the same benchmark.

Google said the capability will improve natural language features across its data services, with BigQuery the implied near-term destination. No API, model card, or technical paper for Gemini-SQL2 has been published as of June 14, meaning the benchmark figure cannot be independently audited or reproduced. The research team has not indicated whether Gemini-SQL2 is a fine-tuned variant, a prompting technique applied to the base Gemini 3.1 Pro model, or a purpose-built system.

Claude Fable 5 launched June 9 with a 95.0% SWE-bench Verified score and an 80.3% SWE-bench Pro score. The 14.7-point gap is consistent with a pattern seen across the entire frontier: Claude Opus 4.5 drops from 80.9% (Verified) to 45.9% (Pro); Gemini 3.1 Pro drops from 80.6% to 46.1%. On Pro's proprietary commercial set, no model exceeds 47.1% as of June 9, 2026, per Scale's standardized leaderboard. The best standardized single-model Pro score was 59.1% for GPT-5.4 using extended inference.

The gap is not a model deficiency so much as a measurement artifact: Verified tasks were public on GitHub during training, making benchmark performance partly a function of data memorization rather than generalizable reasoning. Pro's held-out commercial tasks were never public, making the score a cleaner signal. Anthropic's own benchmarking shows Fable 5's Pro lead over Opus 4.8 (which scored 69.2%) widening on longer, more complex tasks, which is consistent with genuine capability gains. But the absolute Pro number, not the Verified number, is what operators should use when estimating real-world coding performance.

There is also a tokenizer caveat relevant to cost comparisons: Opus 4.7 and later models, including Fable 5, can produce up to 35% more tokens for the same text than pre-4.7 models, meaning per-token price comparisons understate actual per-task costs relative to earlier Claude generations.

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

Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei at 5:21 PM ET on June 12, 2026, invoking national-security export-control authority to bar any foreign national, whether inside or outside the United States, including Anthropic's own non-citizen employees, from accessing Claude Fable 5 and Claude Mythos 5. Because Anthropic cannot filter users by nationality in real time across its cloud delivery infrastructure, it disabled both models for all customers globally within hours. Other Claude models, including Opus 4.8, were not affected.

According to Axios, the administration acted after a separate company claimed it had successfully jailbroken the Mythos model. The administration had previously pressed Anthropic to pause the release of the Mythos-class models and was declined, per the same reporting. Anthropic's public statement said it received only "verbal evidence of a potential narrow, non-universal jailbreak" and characterized the directive as "a misunderstanding," pledging to restore access as soon as possible. As of June 14, no restoration date has been published. The Commerce directive separately requires a license for any export, re-export, or domestic transfer of the two models, placing them in a category previously reserved for advanced chips and weapons-adjacent technologies.

The action has a political context. Anthropic has been on a Pentagon supply-chain-risk blacklist since March 2026 after the company declined to allow military use of Claude for lethal autonomous weapons without safety restrictions. The same week the export order arrived, President Trump had signed a June 2 executive order creating a voluntary pre-deployment testing regime for frontier AI models. The Commerce action, which was not voluntary, effectively superseded that framework for the Mythos class.

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

On May 14, 2026, Colorado Governor Jared Polis signed SB 26-189, which repeals and replaces the original Colorado AI Act (SB 24-205, enacted 2024) and delays the effective date from June 30, 2026, to January 1, 2027. The original law had established a risk-based framework modeled loosely on the EU AI Act, covering AI systems used in consequential decisions affecting employment, housing, healthcare, education, and financial services. SB 26-189 significantly narrows those requirements, shifting toward a transparency-and-notice model with fewer affirmative compliance obligations for developers.

The political path was complicated. A federal magistrate judge had already stayed enforcement of the original law on April 27, 2026. xAI filed a constitutional challenge on April 9. The White House's June 2 executive order specifically named Colorado's AI Act as an example of regulation that, in the administration's view, would compel AI systems to produce distorted outputs to avoid differential impact, and the administration created an AI Litigation Task Force to challenge state AI laws in federal court on Commerce Clause and preemption grounds.

The broader signal: several states that passed comprehensive AI bills between 2024 and early 2026 have now amended, delayed, or substantially narrowed them in response to industry pressure and federal preemption risk. The EU compliance model, with its detailed risk tiers and impact-assessment requirements, has not translated into US law even in states that initially adopted its structure. The Colorado amendment is the starkest single example of that divergence.