AI Export Controls, Legal Wins, and New Hardware: Your Daily Briefing

Anthropic’s Mythos AI Reportedly Breached NSA Systems, Triggering Unprecedented Export Ban

During a June 11 briefing, Senator Mark Warner quoted General Joshua Rudd, head of the NSA and US Cyber Command, stating that Anthropic’s Mythos model penetrated nearly all of the NSA’s classified systems in hours during an authorized red-team exercise. The testimony remains unconfirmed by any government agency, and full classified details are not public. On June 12, the Trump administration directed Anthropic to restrict Fable 5 and Mythos 5 to US citizens only. Unable to verify nationality at scale, Anthropic enacted a full global shutdown, cutting off allies, researchers, and its own foreign-national employees with 90 minutes’ notice. This is the first time the US government has applied export controls directly to an AI model. Anthropic announced Mythos in April, describing it as too capable at finding security vulnerabilities for public release. Access was opened through Project Glasswing to roughly 200 vetted partners including major tech companies and financial institutions. Mythos had uncovered thousands of real-world vulnerabilities, including a 27-year-old flaw in OpenBSD and 271 new bugs in Mozilla’s Firefox 150 browser. Fable 5, released to the public on June 9, uses the same underlying model with safety classifiers. The official trigger for the ban was a jailbreak method for circumventing Fable 5’s safety classifiers, reported to the Commerce Department by Amazon. A researcher using the pseudonym „Pliny the Liberator” published what he claimed was Fable 5’s full system prompt within 48 hours of launch. Anthropic said the jailbreak was narrow and non-universal, a view shared by former Facebook chief security officer Alex Stamos. Trump adviser David Sacks called the jailbreak a „cyber weapon.” The political backdrop includes a February order for federal agencies to stop using Anthropic’s models, a Pentagon designation of Anthropic as a supply chain risk, and a June 2 executive order asking for pre-release access that Fable 5 did not comply with. The ban cut off Five Eyes partners and locked out the UK AI Security Institute. A rare joint statement from the intelligence agencies of all five nations warned that frontier AI models will transform cyber capabilities „not in years, it is months.” Cybersecurity experts noted that the capabilities cited in the Amazon threat intelligence report could already be replicated using older models and open-source Chinese models. As of Monday, Fable 5 and Mythos 5 remained offline. Prediction markets priced 57% odds of restoration before July 1. Trump told Axios that he no longer views Anthropic as a national security threat. An Anthropic identity verification policy set to take effect July 8 could allow restoration domestically. The NSA breach claim has not been independently confirmed.

Anthropic’s Mythos AI reportedly cracked NSA classified systems in hours, that would explain the ban →

AI Law Firm Garfield AI Wins First Court Trial in UK Small Claims Case

Garfield AI, the world’s first AI law firm authorised and regulated by the Solicitors Regulation Authority (SRA), won its first court trial in a small claim brought by a freelancer against a hospitality business. The case involved Tamires Camal Taquidir, who used the platform to pursue £7,000 in unpaid HR service fees. A three-hour trial at Wandsworth County Court in May 2026 resulted in a ruling entirely in favour of the claimant, with the defendant’s counterclaim dismissed. Garfield AI was first authorised by the SRA to provide legal services through AI in May 2025. The firm specialises in English small-claims debt recovery, charging as little as the cost of a cup of coffee. It was founded by lawyer Philip Young and quantum physicist Daniel Long. For the case, Garfield AI handled all pre-trial work, including generating pre-action correspondence, issuing court proceedings, producing documents, and preparing four witness statements and trial bundles. Junior barrister Dominic Li of One Essex Court handled the advocacy in court. Philip Young called the result „a landmark moment for access to justice,” adding that AI made the process „more accessible, more efficient and more affordable” without replacing the judge, barrister, or legal system. Daniel Long said the case showed „what legal AI can do in the real world.” Garfield AI has seen more than 600 claims initiated on its platform and, in just over one year, recovered and resolved over £500,000 for users, with individual claim values ranging from £30 to £10,000.

‘Landmark moment’ - AI law firm wins its first-ever court battle →

Intel and AMD Introduce ACE Specification to Boost AI Performance on x86 CPUs

Intel and AMD have introduced a new x86 specification called Advanced Compute Extensions (ACE) aimed at improving AI operations on CPUs. ACE is designed for smaller models, latency-sensitive tasks, and systems where a GPU is unavailable or not worth the overhead. The specification is built around matrix multiplication, the core of most AI operations. ACE retains the existing AVX10 register structure but adds dedicated hardware for matrix operations, using 512-bit inputs to minimize changes to existing software and hardware workflows. At the instruction level, ACE can carry out up to sixteen times more operations than AVX10 for a given set of input vectors, though real-world performance gains depend on other factors. Power efficiency is highlighted as a practical benefit, especially for edge use cases or single-user applications, since keeping work on the CPU avoids data movement overhead between CPU and GPU. ACE is meant to be implementation-agnostic, providing a consistent target for frameworks like PyTorch and TensorFlow. The extensions support a range of data types: INT8, INT32, FP8, FP16, FP32, and BF16, and include native support for Open Compute Project MX block-scaled formats, which are not part of AVX10. ACE also offers a way to keep certain tasks on the CPU instead of relying on non-standardized NPUs for heterogeneous computing. The specification does not change the role of GPUs in large-scale AI training.

Intel and AMD unveil new x86 standard to make CPUs better at running AI models →

OpenAI to Retire GPT-4.5 and o3 Models Amid User Backlash

OpenAI is retiring GPT-4.5 and o3. The announcement was made at the end of the May 28 release notes for ChatGPT. GPT-4.5 will be removed after a 30-day sunset period ending on June 27, 2026. The o3 model will be removed after a 90-day sunset period ending on August 26, 2026. Both models were previously available only to paying Pro subscribers following backlash after the release of GPT-5. OpenAI stated it is retiring „older models with limited usage in ChatGPT so we can better serve our newer, most capable models.” Users have reacted angrily on Reddit. One Reddit post called the retirement of „the last two good models” and wrote „F*** you, OAI.” Another user said „Thank you so much for letting people know about this horrible news OpenAI buries. I am furious. They just gave me the rest of the reason to cancel my Pro account. There is nothing left. I’ll find o3 on the API, but 4.5 is irreplaceable.” A third user argued „o3 was actually the only model that didn’t have the same guardrails as the rest, and of course they’re gonna remove that?”

OpenAI is officially killing off GPT-4.5 this week in what fans call the ‘end of an era →

Sakana AI Launches Fugu, a Multi-LLM Orchestrator Matching Anthropic’s Top Models

Sakana AI, a Tokyo-based startup, launched Fugu, a multi-LLM orchestrator that dynamically coordinates multiple AI models from a swappable agent pool. Fugu is itself a language model trained to call other LLMs, including copies of itself, and either handles a task alone or assembles a team of specialized models. Users access the system through a single OpenAI-compatible API. Two variants were introduced: the base Fugu model for low latency and everyday performance, and Fugu Ultra for maximum answer quality on complex, multi-step problems. Early users applied Fugu Ultra to AI research, reproducing scientific papers, cybersecurity analysis, and patent and literature searches. According to benchmark results published by Sakana AI, Fugu Ultra performs on par with Anthropic’s Fable 5 and Mythos Preview across coding, reasoning, science, and agent benchmarks. Neither Anthropic model is in Fugu’s agent pool because they are not publicly available. A detailed benchmark table shows Fugu Ultra scoring competitively against Opus 4.8, Gemini 3.1 Pro, and GPT 5.5 on twelve benchmarks, including SWE Bench Pro (73.7), TerminalBench 2.1 (82.1), and LiveCodeBench Pro (90.8). Sakana AI pitched Fugu as a safeguard against single-provider dependence, citing the export controls on Anthropic’s models as an example. The company stated that relying on a single company’s APIs for critical infrastructure is a material vulnerability. Fugu’s model pool is fully swappable, so the system can reroute to other models if one provider goes dark. About 500 beta users tested the system; one software developer reported that Fugu Ultra caught more than twenty bugs during code review. Both variants are live through a single API with subscription and usage-based billing. Fugu’s technical approach builds on Sakana AI’s research papers Trinity and Conductor, presented at ICLR 2026. Sakana AI was founded by former Google AI researchers Llion Jones and David Ha; Jones co-authored the seminal 2017 paper „Attention Is All You Need” that introduced the Transformer architecture.

Sakana AI’s Fugu orchestrates multiple LLMs to match Anthropic’s Fable and Mythos benchmarks →