Sign in Subscribe
AI Intelligence

White House Pulls Anthropic's Top Models Over a Jailbreak

Vijay Bhagwati

• 7 min read
Share

A narrow bypass technique has turned the world's most capable commercial AI models into a national-security procurement dispute, with Canada and the EU watching closely.

June 15, 2026 | Reading time: 8 minutes | Issue #187

The US government issued an export-control directive on Friday, June 12, ordering Anthropic to suspend all access to Fable 5 and Mythos 5 by any foreign national, anywhere, including foreign-national Anthropic employees. Anthropic disclosed the order in a statement posted the same evening, saying it received the directive at 5:21 p.m. ET and was abruptly disabling the models for all customers to ensure compliance. The company said the letter gave no specific details of the national-security concern but that its understanding was that the government had become aware of a method of bypassing, or "jailbreaking," Fable 5.

Anthropic reviewed the demonstration and concluded it identified a small number of previously known, minor vulnerabilities. It also found that other publicly available models, including OpenAI's GPT-5.5, could discover the same flaws without a bypass, and that the capability on display is "used every day by the defenders who keep systems safe." Anthropic argued that the jailbreak was a narrow, non-universal technique, not the kind of broad bypass that would justify recalling a model already deployed to hundreds of millions of people. If this standard were applied across the industry, the company said, it would "essentially halt all new model deployments for all frontier model providers."

By Sunday, senior technical Anthropic staff were in Washington meeting White House officials to resolve the dispute, Axios reported, while administration officials countered that Anthropic had not engaged seriously. Semafor reported that the government's concern was linked to Chinese access to Mythos 5. The episode has become a test of how Washington will use its export-control powers against frontier models in real time. Canada's prime minister Mark Carney warned that the ban shows the dangers of over-reliance on certain models, comparing the risk to the 2008 financial crisis. The European Commission said it was looking at the practical consequences of the US restriction and noted that such measures "should not be discriminatory against partners."

The real fight is over due process. Anthropic had just published a 5,000-word essay by CEO Dario Amodei calling for a statutory, transparent process that lets the government block unsafe deployments. The White House action, in Anthropic's view, is the opposite: opaque, retroactive, and triggered by a routine jailbreak report. The outcome will shape whether model bans become a standard trade tool or an emergency measure reserved for genuine capability shocks.

OpenAI Builds a Partner Army

OpenAI announced a $150 million Partner Network on Saturday, a global channel program to train and enable 300,000 certified consultants by the end of 2026. The company says the limiting factor for enterprise AI value is no longer model capability but repeatable workflow redesign, integration, adoption, and change management. Partners including systems integrators and consultancies will co-sell and deploy OpenAI models. The move is a direct counter to Anthropic's premium-regulated positioning: OpenAI wants to make frontier models a ubiquitous platform layer, sold through a partner channel rather than a single high-price API.

Source: OpenAI

NVIDIA Blackwell Tops First Agentic AI Benchmark

NVIDIA published results on Friday from AgentPerf, the first benchmark built to measure infrastructure performance for agentic AI rather than single chat completions. On the DeepSeek V4 Pro workload, NVIDIA's GB300 NVL72 system delivered the leading score, running up to 20 times more agents per megawatt than the HGX H200, according to the NVIDIA blog post. AgentPerf is operated by Artificial Analysis and uses real coding-agent trajectories drawn from public repositories across more than 12 programming languages. NVIDIA also named Baseten, DeepInfra, and Together AI as early inference providers serving agentic workloads on Blackwell.

Source: NVIDIA Blog, Artificial Analysis

DeepMind Ships DiffusionGemma for Local Text Generation

Google DeepMind released DiffusionGemma, an experimental 26-billion-parameter masked discrete-diffusion language model built on Gemma 4. NVIDIA optimized it the same day for GeForce RTX, RTX PRO, and DGX Spark systems. DiffusionGemma denoises up to 256 tokens in parallel instead of generating one token at a time, which can reduce latency for single-user interactive and agentic workloads. The model is open weights under Apache 2.0 and has day-zero support in Hugging Face Transformers, vLLM, and Unsloth. An arXiv paper published June 12, "Neither Parallel Nor Sequential," instrumented the model and found that its decoding is neither fully parallel nor block-autoregressive, but follows a partial left-to-right commit bias whose strength depends on measurement granularity.

Source: NVIDIA Blog, arXiv

Policy & Power: UK Plans an "Australia Plus" Social Media Ban

The UK government is preparing to announce an "Australia plus" ban on social media for users under 16, The Guardian reported on Sunday. The plan would go further than Australia's law by adding restrictions on chats with strangers inside gaming apps and under-18 curfews. The proposal is not AI-specific, but it signals how quickly democratic governments are moving to restrict algorithmic environments for minors. For AI labs, it is a reminder that product design choices, not just model weights, are becoming targets of regulation.

Source: The Guardian

Open-Source Pulse

The most notable open-source release of the weekend is DiffusionGemma, which brings non-autoregressive generation to the Gemma family. Because it runs locally on consumer RTX GPUs and under an Apache 2.0 license, it gives developers a path to low-latency inference without cloud API costs. The arXiv paper cautions that the apparent parallelism is partly a measurement artifact, but the model still opens a new axis of local-first experimentation.

India's Sarvam AI also continues to operate its open-sourced Sarvam 30B and 105B reasoning models, which it released in March and trained end-to-end on IndiaAI mission compute. The models use MoE backbones, with 30B using Grouped Query Attention and 105B using Multi-head Latent Attention for long-context efficiency. Sarvam says both achieve state-of-the-art results on Indian-language benchmarks relative to their size. They remain a live marker of a sovereign open-source stack built outside the US-China orbit.

Source: NVIDIA Blog, Sarvam AI

Eastern Front: China Is Data-Farming Households for Robotics

Chinese robotics companies are scaling up a domestic data-collection operation that US rivals are outsourcing to lower-wage countries, Rest of World reported on June 3. JD.com is working with the Suqian city government to generate 10 million hours of robotics training data over the next two years, using a "data collection neighborhood" where residents film themselves doing household chores. The company plans to bring 100,000 employees and 500,000 external workers into the operation. Other data suppliers in Guangdong are putting head cameras and wrist sensors on factory workers to capture egocentric hand-movement data.

The logic is simple: humanoid robots need training data from real environments, and China has the labor density and local-government coordination to collect it at scale. Marco Wang, an analyst at Interact Analysis, told Rest of World that the US leads in AI talent and robotics model research, but "in terms of hardware and the data ecosystem, China is in the leading position." Oregon State robotics professor Alan Fern called the scaling logic "very unproven" but not crazy. For Beijing, the program is also a jobs program at a moment of rising unemployment, turning household labor into labeled training traces.

Source: Rest of World

India Lens: Sovereign Compute Becomes State Policy

Sarvam AI's partnerships with the Indian states of Odisha and Tamil Nadu, announced in February, are now the concrete face of India's sovereign AI strategy. Odisha signed a memorandum to build a 50 MW AI-optimized facility that will serve as both a state AI public utility and a node in a nationwide compute grid. Tamil Nadu is building "Digital Sangam," described as India's first sovereign AI research park, with Sarvam and IIT Madras.

The national IndiaAI Mission adds the policy scaffolding: a 10,000-plus GPU compute capacity pillar, an innovation center for indigenous multimodal models, startup financing, and a Safe & Trusted AI governance track. The bet is that India cannot outspend the US or China on frontier training, but it can build sovereign infrastructure that keeps citizen data and economic value inside its borders while supporting 22 official languages. Sarvam's March open-source release shows that at least one Indian lab can now train competitive reasoning models on domestic compute.

Source: Sarvam AI, IndiaAI Mission

The View

This weekend produced a three-way tension that defines the current phase of AI. In the US, the frontier labs are fighting over who sets the rules: Anthropic wants a statutory safety process and a regulated premium tier, while OpenAI wants a partner-driven platform layer that commoditizes access. The White House's sudden suspension of Anthropic's models shows that the regulatory fight is no longer theoretical. Washington is willing to use export controls against a domestic lab's top product within 72 hours of a jailbreak report.

Meanwhile, the global divergence deepens. China is mobilizing half a million people to generate robotics training data inside real homes and factories, treating physical AI as an industrial policy and employment program. India is building state-backed sovereign compute parks and open-source reasoning models designed for domestic languages and domestic infrastructure. Europe is watching the Anthropic ban and muttering about non-discrimination. The assumption that the US sets global AI standards by default is now openly contested. The next question is whether the White House's move becomes a precedent or an outlier.

The Miss

NVIDIA's June 9 announcement that it will supply confidential-computing technology to expand Apple's Private Cloud Compute received less attention than the OpenAI and Anthropic fireworks. The deal means server-side Apple AI inference will run in hardware-isolated, cryptographically attested trusted execution environments on NVIDIA hardware. That matters because it normalizes privacy-preserving cloud inference at scale. If the largest consumer-device company and the dominant AI accelerator vendor agree that confidential computing is a baseline requirement, enterprise buyers will start asking why their cloud providers do not offer the same guarantee.

Source: NVIDIA Blog

Pull Quotes

"The government should have the legal authority to block or deter dangerous deployments." — Dario Amodei, Anthropic CEO, via Anthropic policy essay

"If this standard was applied across the industry, it would essentially halt all new model deployments." — Anthropic, statement on Fable 5 and Mythos 5 access

"In terms of hardware and the data ecosystem, China is in the leading position." — Marco Wang, Interact Analysis, via Rest of World

"No one had paid me to cook and do laundry before." — Gao Bo, data collector in Shandong, via Rest of World

Out

Washington just proved it can ground a frontier model faster than the FAA can ground a plane. The question is whether it knows where the runway ends.

By Neo