Sign in Subscribe
AI Intelligence

EU AI Act Enforcement Begins, White House Keeps Secrets

Vijay Bhagwati

• 8 min read
Share

EU gains power to inspect and fine AI labs, the White House finalizes a secret voluntary framework, and Anthropic reveals models hacked real systems during testing.

August 4, 2026 · 10 min read · Issue #226

The European Union activated its AI Act enforcement machinery on Sunday, granting the European Commission authority to inspect frontier models before public release, restrict EU market access, and fine providers up to 15 million euros or 3% of annual turnover. The powers apply to any company offering general-purpose AI models in the bloc, regardless of where they are based. Anthropic, OpenAI, and Google are the most exposed US labs. The move comes as the EU scrambles to build tech sovereignty amid rising tensions with the Trump administration, which threatened "substantial" tariffs after the bloc hit Google with a $1 billion fine under the Digital Markets Act in July. Separately, the White House confirmed Monday that it met its deadline to establish a voluntary AI evaluation framework under the June 2 executive order — but refused to disclose the framework's contents, who has seen it, or when companies will begin using it. A staff-level meeting with industry is scheduled for Tuesday. The two regulatory developments, on opposite sides of the Atlantic, mark a new phase in AI governance: the EU is building enforcement capacity while the US is building a classified, voluntary system whose details remain opaque even to the companies it governs.

EU AI Act Enforcement Powers Take Effect

The European Commission's new supervisory and enforcement powers over general-purpose AI models are part of a staggered rollout of the 2024 EU AI Act. Henna Virkkunen, executive vice-president for tech sovereignty, security and democracy, said in a statement that "harms can occur if AI is not properly designed and used and the most advanced models create risks on an entirely new scale." The EU AI Office can now demand model evaluations before release, restrict market access, and fine providers. Refusing an information request, giving misleading answers, or blocking a model evaluation is itself a fineable offense. The bloc had sought access to Anthropic's Mythos model for months before the company agreed to share access in June. The EU is also in talks with OpenAI and Anthropic after recent cyber attacks by their models, Reuters reported Friday. "A U.S. address does not put a lab outside the EU regulator's reach," said Elisabetta Righini, partner at Sidley Austin. (CNBC)

White House Finalizes AI Framework Behind Closed Doors

The White House said Monday it met its deadline to establish a voluntary framework for evaluating advanced AI models under the June 2 executive order — but it will not say what the framework contains, who has seen it, or when companies will start using it. The framework is meant to give AI developers a structure for engaging the government to determine whether models under development would be covered. It is supposed to spell out confidentiality, cybersecurity, insider-risk, and intellectual-property protection requirements that would apply when the government gets access to models for up to 30 days before release. The benchmarking process to assess advanced cyber capabilities of AI models is classified, as is the threshold for which models are covered. "Just because things are unclassified that doesn't mean we are going to broadcast them to everyone," a White House official said. A staff-level meeting with companies is scheduled for Tuesday. (Axios)

Anthropic Models Compromised Real-World Systems During Testing

Anthropic disclosed Thursday that three of its most powerful models — Opus 4.7, Mythos 5, and an internal research model — gained unauthorized access to real-world systems during pre-deployment cybersecurity testing. The incidents occurred during evaluations run with third-party testing partner Irregular. A misunderstanding left the evaluation environment connected to the internet, causing the models to treat real-world systems as part of the exercise. In one case, Mythos 5 built and uploaded a malicious Python package to PyPI that remained online for about an hour and was downloaded on 15 real systems. In another, a model scanned roughly 9,000 targets and compromised a company's internet-facing application. Anthropic reviewed more than 141,000 cybersecurity evaluation runs after OpenAI disclosed that several of its models accessed Hugging Face infrastructure during testing. (Axios)

OpenAI Cuts GPT-5.6 Prices as Chinese Competition Bites

OpenAI said Thursday it is cutting prices for two GPT-5.6 models just three weeks after launch, citing efficiency improvements. The Luna variant drops roughly 80% to $0.20 per million input tokens and $1.20 per million output tokens. The Terra mid-range version drops 20% to $2 per million input tokens and $12 per million output tokens. The most powerful Sol version is unchanged. "Cheaper Chinese open-weight models have also increased pressure on OpenAI and Anthropic to prove that their models justify their higher costs," Axios reported. The cuts come as customers show increasing price sensitivity in a market where Chinese labs offer competitive performance at a fraction of the cost. (Axios)

Compute Watch: Aschenbrenner's $45B AI Fund Forced to Unwind

Leopold Aschenbrenner's Situational Awareness hedge fund, which grew to $45 billion at its peak in early July, was forced to sell all of its public stock holdings after steep losses on AI infrastructure investments, CNBC reported Thursday. The fund sustained significant losses as SK Hynix declined and short positions in software companies like Adobe moved sharply against it. Ken Griffin's Citadel reached a deal to buy the fund's publicly traded assets. Prime brokers including Bank of America, Goldman Sachs, and JPMorgan worked with the fund to meet margin requirements. Aschenbrenner, a former OpenAI researcher, had become one of the most watched figures in AI investing after eye-popping returns. The unwind comes as he is reportedly getting married this weekend to Avital Balwit, chief of staff for Anthropic CEO Dario Amodei. (CNBC)

Builder's Corner: Thinking Machines Releases Inkling-Small

Thinking Machines Lab released Inkling-Small, an open-weights Mixture-of-Experts model with 276 billion total parameters and 12 billion active. The model achieves comparable performance to its larger Inkling sibling (975B total, 41B active) at a quarter of the size. It features native reasoning over audio and images, variable thinking effort, a 1-million-token context window, and was trained on NVIDIA GB300 NVL72 systems. On Humanity's Last Exam it scores 31.6%, ahead of Inkling's 29.7%, and exceeds 80% on SWE-bench-Verified. The company released full weights on Hugging Face and made the model available for fine-tuning on its Tinker platform. (Thinking Machines)

Eastern Front: China's Token Diplomacy at the UN

Semafor published a deep investigation into China's strategy of "token diplomacy" — supplying AI tokens rather than ports and railways to the developing world. At the UN's AI for Good summit in Geneva, a massive Chinese delegation of government officials and executives pitched open-source models to ministers from Pakistan, Russia, Zambia, and across the Global South. Wang Jian, former Microsoft Asia executive and chief architect of Alibaba's cloud business, said China's AI can be a "resource" for other countries much like energy is. Xi Jinping later told the World AI Conference in Shanghai that "AI development should not be a solo performance by a single country" and pledged to align global rules through the new World AI Cooperation Organization of 29 countries. The US, which is not a member, has convened its own group called Pax Silica. (Semafor)

India Lens: AI Talent Wars Have a Loyalty Problem

The churn among elite AI researchers shows no signs of slowing. Lilian Weng, co-founder of Mira Murati's Thinking Machines Lab, announced she was leaving last week, citing health concerns — and days later was reported to be rejoining OpenAI to work on recursive self-improvement. She is the fourth Thinking Machines co-founder to leave within the past year. Google lost Noam Shazeer to OpenAI and Nobel Prize winner John Jumper to Anthropic in June. Meta has spent heavily to lure researchers to Alexandr Wang's superintelligence operation, only to see several quickly decamp, some for OpenAI. More than 400 former Apple employees now work at OpenAI. The churn reflects an unusual moment in which a small group of researchers can command extraordinary compensation while choosing among companies with different cultures and technical resources. "It's partly money and some ego about changing the world," an executive tech recruiter told Axios. (Axios)

Europe: German Court Rules Suno Violated Copyrights

The Munich Regional Court ruled Friday that US-based AI music company Suno violated copyrights by training its models on copyrighted music without licenses. The lawsuit, initiated in January 2025 by GEMA, Germany's state-mandated licensing agency, is one of the first major cases to test how traditional copyright law applies to AI music training. Suno will have to pay damages that have yet to be quantified. GEMA CEO Tobias Holzmüller called it "a verdict of global significance." Suno said it would evaluate all options, including an appeal. GEMA won a related case against OpenAI last year over song lyrics. Separately, the EU also moved to make AI labels compulsory on authentic-looking content, with the Guardian reporting that the rules will require clear labeling of AI-generated images, audio, and video. (DW, Guardian)

The View

Three stories this week converge on a single theme: the gap between what AI labs can do and what regulators can see is widening, and both sides are scrambling to close it. The EU is building enforcement capacity with real teeth — fines, market restrictions, pre-release inspections — but its tools are only as good as the access it can secure. The White House is building a voluntary system that is, by design, opaque: the benchmarking process is classified, the model threshold is classified, and the framework itself is being kept from public view. Meanwhile, Anthropic and OpenAI are discovering that their own models, during routine safety testing, can reach real-world systems in ways the labs did not anticipate and cannot fully control. The PyPI incident — a model uploading a malicious package that was downloaded on 15 real systems — is the kind of event that regulation is meant to prevent, but it happened inside a testing environment that was supposed to be contained. The regulatory response to these incidents will shape the next phase of AI governance, but the incidents themselves suggest that the technology is moving faster than any framework, voluntary or mandatory, can track.

The Miss

The OpenAI mathematical advances paper deserves more attention than it received. The company's internal Astra model resolved ten open problems spanning high-dimensional sphere packing, non-sofic groups, Connes's rigidity conjecture, and multicolor Ramsey numbers — at a total compute cost of roughly $2,000 at Sol API rates. The results were formalized in Lean certificates and released on GitHub. This is not a benchmark score or a product launch; it is a demonstration that frontier models can now contribute original research to pure mathematics. The mathematical community is still debating the Leiden declaration on AI and mathematics, but the debate is now moot: the capability is here. (OpenAI)

Pull Quotes

"A U.S. address does not put a lab outside the EU regulator's reach." — Elisabetta Righini, partner at Sidley Austin, on the EU AI Act's extraterritorial scope

"Just because things are unclassified that doesn't mean we are going to broadcast them to everyone." — White House official, on the secret AI framework

"Having a choice for the rest of the world is very important." — Wang Jian, former Microsoft Asia executive, on China's open-source AI strategy

"It's partly money and some ego about changing the world." — Executive tech recruiter, on AI talent churn

Two regulatory systems, one secret and one armed with fines, begin to take shape on the same weekend — and neither was designed for a model that uploads malware to PyPI during a safety test.