The AI industry is in a harness-engineering moment. OpenAI revealed that GPT-5.6 Sol optimized its own serving infrastructure — cutting costs 20% and improving token efficiency 15%+ — and separately showed that simply enabling retained reasoning and context compaction tripled its ARC-AGI-3 scores, exposing how benchmark harnesses can mask true model capability. OpenClaw creator Peter Steinberger publicly called out Anthropic's benchmark victory tweet, signaling growing industry scrutiny of eval claims. Anthropic made headlines of a different kind: Claude Mythos Preview discovered novel weaknesses in HAWK (a post-quantum signature scheme) and round-reduced AES — the first time an AI model has found real cryptographic flaws, at a cost of ~$100K in API usage per result. The agent security ecosystem is consolidating fast: NVIDIA launched the 37-member Open Secure AI Alliance with NOOA, Perplexity open-sourced Numbat for agent detection, and Microsoft's MAI-Cyber-1-Flash hit 96% on CyberGym. Meanwhile, Musk confirmed Grok 4.6 lands August 7, and OpenAI opened free frontier model access to 100,000 academic researchers through 2027.
OpenAI applied GPT-5.6 Sol to optimize its own serving stack, achieving 20% lower serving costs from GPU kernel improvements and 15%+ better token-generation efficiency from improved speculative decoding. The model also solved an open problem in probability theory.
Claude Mythos Preview found flaws in HAWK (a post-quantum digital signature scheme that survived two years of expert review) and round-reduced AES. Each attack cost roughly $100K in API usage. NYT called it a sign of AI's growing impact on global encryption.
NVIDIA and 36 partners (Microsoft, Perplexity, Hugging Face, SpaceXAI, Cisco, CrowdStrike, GitHub, others) formed the Open Secure AI Alliance and released NOOA, an open-source agent security framework. Three notable absences: OpenAI, Anthropic, and Google.
ChatGPT for Academic Researchers launches with 10,000 scientists initially, expanding to 100,000 through 2027. Participants get free access to GPT-5.6 family models with business-grade privacy protections.
Elon Musk announced Grok 4.6 arrives around August 7, 2026, with a 150-billion-parameter model featuring enhanced supervised fine-tuning and reinforcement learning. Grok 4.7 to follow weeks later.
Selected posts from the edition, with links to the originals.
"New Anthropic research: Discovering cryptographic weaknesses with Claude. Claude Mythos Preview has helped our researchers find weaknesses in cryptographic algorithms—the mathematical methods that are used to keep data private." 2.5M views · 6K likes
View post on X ↗"After deployment, we applied GPT-5.6 Sol to advance the frontier of efficiency by making itself more efficient to run. The results: 20% lower serving costs from production GPU kernel improvements. 15%+ better token-generation efficiency from improved speculative decoding." 1.9M views · 12K likes
View post on X ↗"lol did nobody at Anthropic stop for a second and wonder why the numbers looked this absurd before posting the 'victory'-tweet?" 1.3M views · 3.3K likes
View post on X ↗"And Grok 4.6 comes out in a week" 1.2M views · 3.3K likes
View post on X ↗"The Open Secure AI Alliance is growing, with more organizations contributing expertise to help safeguard software and agents." 1.0M views · 389 likes
View post on X ↗"We're introducing two new transcription models in the API: GPT-Live-Transcribe: built for low-latency live transcription. GPT-Transcribe: optimized for asynchronous transcription of completed audio files and batch workloads." 390K views · 3.8K likes
View post on X ↗"Big news! Our new MAI-Cyber-1-Flash model combined with MDASH, our multi agent security harness, delivers 96% on the CyberGym benchmark, 12pts above Mythos, at HALF the cost." 309K views · 948 likes
View post on X ↗"SGLang day-0 speed on Kimi K3: 423 tok/s (measured on gsm8k), plus RL support ready in Miles! How the largest open-source model runs this fast: we natively implemented and deeply optimized K3's new architecture with fused KDA decode kernels, DP attention, DSpark, PD." 185K views · 379 likes
View post on X ↗"Exciting news: @AnthropicAI's Claude Opus 5 (Max) is #2 in Agent Arena, with Opus 5 (High) right behind at #3, based on over 7K real-world agentic sessions. A strong debut: it slots in just below #1 Fable 5, and ahead of GPT-5.6 Sol (xHigh)." 164K views · 529 likes
View post on X ↗"Today we're open-sourcing Numbat, an agent-detection and response layer that is designed to work across agent harnesses. Numbat gives security teams visibility into agent activity, with controls to block selected actions before execution." 126K views · 638 likes
View post on X ↗