Anthropic's top models blocked abroad; AI governance expands globally
US export curbs on Anthropic's top models draw pushback as states, EU, and global regulators advance AI rules.
By BINA Editorial
Today's brief includes a follow-up on the Anthropic model suspension that dominated Friday's news.
US bars foreign access to Anthropic's newest AI models
On 13 June, the Trump administration ordered Anthropic to suspend global access to its two most powerful models—Fable 5 and Mythos 5—citing a national-security concern over a reported jailbreak vulnerability. Because Anthropic cannot reliably verify user nationality at login, it complied by taking both models offline for all customers, including its own foreign employees. The company publicly disputed the restriction's breadth, stating that the government provided only "verbal evidence of a potential narrow, non-universal jailbreak." The episode prompted Canadian Prime Minister Mark Carney to warn of the dangers of over-relying on a small number of US AI providers, as Canada accelerates efforts to diversify its technology base. (Follow-up; story first reported 13 June.)
US states advance AI legislation despite federal preemption pressure
Despite a White House executive order threatening legal challenges and federal funding cuts against states that regulate AI, a growing number of US states are pressing ahead. Connecticut, Washington, and Utah have enacted laws requiring digital content to carry markers indicating AI origin. Colorado now mandates that companies using AI to make decisions in employment, housing, banking, or education must inform the affected individuals. California's legislature is considering a "No Robo Bosses Act of 2026" that would bar employers from using AI as the sole arbiter of disciplinary actions or dismissals. New York lawmakers passed bills covering chatbot safety for children, AI training-data transparency, and a ban on algorithmic retail-pricing.
EU finalises guidance on classifying high-risk AI systems
The European Commission published draft non-binding guidelines clarifying how companies should classify "high-risk" AI systems under the EU AI Act. The guidelines state that complex agentic systems—composed of multiple AI components—must be assessed holistically rather than piece by piece, and they include worked examples spanning healthcare, education, and public administration. A public consultation is open until 23 June 2026; the Commission will finalise the guidance in time for the Act's high-risk provisions to take effect.
Financial Stability Board proposes AI governance standards for banks and insurers
On 10 June, the Financial Stability Board released a consultation report proposing twelve sound practices for financial institutions adopting AI responsibly. The framework covers governance structures, model risk management, third-party dependencies, and data quality, targeting institutions that are integrating AI into credit scoring, fraud detection, and trading systems. The FSB is accepting feedback until 22 July 2026, after which it plans to issue final guidance for national regulators worldwide.
Japanese researchers develop interpretable AI for materials science
A team at the Institute of Science Tokyo published a method on 15 June for making AI models used in materials discovery interpretable. By analysing the features a trained model extracts from crystal structures and optical spectra, the technique groups materials with similar structural and optical characteristics, revealing hidden relationships that conventional laboratory searches would miss. The researchers say the approach can guide the design of better solar cells, LEDs, and other optical materials by telling scientists precisely where to focus their experiments.
Claude Opus 4.8 displaces GPT-5.5 at the top of AI benchmarks
Anthropic's Claude Opus 4.8, released in late May, has taken the top position on the Artificial Analysis leaderboard, registering a 61.4% blended score and a leading 1,545 Elo rating, displacing OpenAI's GPT-5.5. The model excels in reasoning and agentic tasks, including an 88.6% score on SWE-Bench Verified, a standard measure of autonomous software-engineering ability. GPT-5.5 retains a narrower edge in dedicated coding benchmarks, while Chinese models Qwen 3.7 Max and MiniMax M3 are matching Western flagship performance at significantly lower cost.