Mythos cracks crypto, Amazon pivots, and the open-weights war heats up

Anthropic’s Mythos model discovers cryptographic weaknesses in HAWK and AES

Anthropic’s AI model Claude Mythos Preview has identified mathematical vulnerabilities in cryptographic algorithms, including an improved attack on the post-quantum signature scheme HAWK and a new attack on a reduced version of the Advanced Encryption Standard (AES). Neither finding affects systems in use today: HAWK is only a candidate in an ongoing standardization process by NIST, and the AES attack applies to a modified version using 7 of 10 rounds. Mythos found the HAWK attack in 60 hours for $100,000 in API costs, exploiting a previously undetected symmetry in the mathematical lattice. The model worked semi-autonomously in a multi-agent system, with one agent initially dismissing the idea and a second finding a way to exploit it. The human researcher, who was not an expert in lattice-based cryptography, acted mostly as project manager.

Anthropic says its Mythos model found vulnerabilities in cryptographic algorithms that secure the internet →

Amazon scales back Nova AI models, shifts focus to Frontier research team

Amazon is overhauling its AI strategy, scaling back most in-house Nova models including the flagship Nova Premier and Omni models, the Reel video model, and the Canvas image generator, according to Business Insider. Those models remain online in “keep the lights on” mode for existing customers but are no longer actively developed. The pullback follows layoffs in the AGI division and closure of the AGI Lab founded in 2024. Resources now flow to the Frontier Model Research group under Pieter Abbeel, who joined through the acquisition of robotics startup Covariant. A new foundation model is expected at the re:Invent conference this fall. Amazon remains one of the biggest backers of Anthropic and OpenAI, where investments could top $100 billion.

Amazon reportedly scales back its Nova AI models and bets on a new Frontier research team →

Anthropic denies open-weights ban calls as 77 firms sign letter supporting open models

Anthropic CEO Dario Amodei published a post on July 27, 2026, stating the company has never called for a prohibition on open-weights AI models, responding to a letter signed by 77 organizations including Nvidia, Google, Meta, Microsoft, OpenAI, IBM, and SpaceX. The letter argues that open weights expand access to the AI economy and strengthen competition, while acknowledging that released weights cannot be withdrawn. Amodei’s post accepts open-weights models without dangerous capabilities as a public good but warns that sufficiently capable open models may present higher risks for cyberattacks or biological attacks. Anthropic supports three alternative measures: refusing to sell powerful chips to China, targeting industrial-scale distillation, and mandatory pre-release safety testing for all sufficiently capable models. The post notes that the Commerce Department last year considered adding Chinese AI labs to its Entity List, but those efforts were killed by officials concerned about stifling innovation.

Anthropic faces open-weights ban accusations as 77 firms sign letter →

Nvidia invests in Ilya Sutskever’s SSI lab, shifting away from Google chips

Nvidia is investing a substantial but undisclosed sum into Safe Superintelligence (SSI), the AI lab led by Ilya Sutskever, OpenAI’s former chief scientist. As part of a long-term partnership, SSI gains access to Nvidia’s next-generation Vera Rubin GPU platform, which the companies say will boost its compute capacity tenfold. Until now, SSI had relied mainly on Google’s TPU chips. The deal also helps Nvidia fend off Google as a rival chip supplier. SSI was founded in 2024 with a single goal: a straight-shot sprint toward safe superintelligence. The company raised about $2 billion from Andreessen Horowitz, Sequoia Capital, DST Global, and Greenoaks, and was most recently valued at roughly $30 billion.

Nvidia invests in Ilya Sutskever’s AI lab, shifting SSI away from Google chips →

Moonshot AI releases Kimi K3, the largest open-source model with 2.88 trillion parameters

Moonshot AI has released Kimi K3, the largest open-source AI model to date, with 2.88 trillion parameters. It uses a mixture of experts (MoE) architecture that dynamically allocates computational resources, achieving 2.5 times the intelligence per compute unit compared to earlier systems. The model features a 1-million-token context window and native multimodal capabilities for integrating text and visual data. Moonshot AI provides open-source tools including high-performance attention kernels, communication libraries, and sandbox environments. The licensing framework offers free access for individuals and small organizations, while commercial entities with over $20 million in annual revenue or more than 100 million users must obtain a commercial license. In performance benchmarks, Kimi K3 rivals proprietary models such as Fable 5 Max and GPT 5.6 Soulmax.

2.88 Trillion Parameter Kimi K3 Largest Open Model is Now Available →