1. OpenAI Ships GPT-5.6-Cyber, Its First Model to Cross the 'High' Cyber Threshold
OpenAI released GPT-5.6-Cyber on August 10 through a gated program called Daybreak Red, and assessed it as reaching High cybersecurity capability under its Preparedness Framework — the first model to do so — while staying below the Critical threshold. The model completed 95% of requests involving exploit chains, authentication bypass, and privilege escalation, against just 1.5% for the standard guardrailed GPT-5.6. Access is limited to vetted defenders with identity verification and monitoring, and every individual Daybreak account must use a hardware security key starting September 1.
2. An AI Model Found Two Unknown Chrome Zero-Days — Google Has Patched Them
Pointed at V8, the JavaScript engine inside Chrome, GPT-5.6-Cyber surfaced two previously unknown vulnerabilities that could be chained to corrupt memory and escape V8's heap sandbox. The bug targets TurboFan, V8's optimizing compiler, which collects runtime type information and uses it to strip safety checks it judges unnecessary. Google shipped a fix as CVE-2026-15903, affecting Chrome builds before 150.0.7871.128. It lands amid a broader surge in AI-assisted bug hunting that has pushed Chrome patch volume sharply higher.
3. Meta Returns to Open Source with Muse Glimmer, a 30B Apache 2.0 Agentic Model
Meta open-sourced Muse Glimmer, a 30-billion-parameter multimodal model distilled from Muse Spark and tuned for always-on local agent work — function calling, local coding, and LLM-as-a-judge evaluation. It ships under Apache 2.0, a more permissive license than Llama ever carried, and is the company's first fully open release since it replaced the open-weight Llama family with the proprietary Muse Spark in April. It runs on a single consumer GPU with 24GB of VRAM, with day-0 support in transformers, llama.cpp, and vLLM.
4. The EU Started Actually Enforcing the AI Act on August 2
The European Commission's AI Office and national authorities began enforcing the AI Act, with Article 5 (prohibited practices) and Article 50 (transparency) now live. Chatbots must identify themselves as automated, deepfakes must be labelled, and machine-generated or edited content must carry machine-readable marks. Penalties run from €7.5 million to €35 million, or 1% to 7% of global turnover. The Commission also published a first list of 180+ organizations that signed the Code of Practice on AI-generated content transparency. Obligations for high-risk systems were pushed back under the Digital Omnibus amendments.
5. Cognition in Talks at a $40 Billion Valuation, Up 50% in Under Three Months
The coding-agent startup behind Devin is in early discussions with investors on a round that would lift its valuation past $40 billion, Bloomberg reported August 12. That is a jump of more than 50% from the $26 billion valuation at which it raised $1 billion less than three months ago. It fits a wider pattern: capital keeps concentrating in US AI labs, compute, and regulated verticals, with Harvey reported around $8 billion for legal AI and Chai Discovery raising $400 million at $3.8 billion.
6. Anthropic Makes Claude Sonnet 5's Introductory Pricing Permanent
As of August 10, Sonnet 5's launch pricing of $2 per million input tokens and $10 per million output tokens is permanent rather than promotional. Released June 30, Sonnet 5 is the most agentic model in the Sonnet tier so far, closing much of the gap with Opus 4.8 on agentic coding and computer use — and outscoring it on knowledge work — at a fraction of the cost. Opus-class models still hold the edge on the highest-accuracy tasks, but the price-performance floor for agent workloads keeps dropping.
7. Nature Publishes 'The AI Scientist': End-to-End Automation of AI Research
Sakana AI's system, now peer-reviewed in Nature, takes a broad research direction and runs the full loop autonomously — generating novel ideas, reading the literature, designing and coding experiments via parallelized agentic tree search, and writing the paper. One of its outputs passed the first round of peer review at a workshop of a major machine learning conference. A separate Nature study this year used a pretrained language model to identify AI-augmented research across 41 million natural-science papers, giving the first real measurement of how far this has already spread.
8. Nvidia's Datacenter Roadmap Bends Toward Agents, Not Just Training
Nvidia's Vera CPU — 88 custom Olympus cores, 176 threads, 1.2 TB/s of bandwidth and 164MB of L3 cache — is pitched squarely at the bottlenecks in agentic AI and reinforcement learning rather than raw pretraining throughput. Inside the liquid-cooled Vera Rubin NVL72 rack platform, Nvidia claims local agent runtime loops complete up to 1.8x faster than on traditional x86 infrastructure, with a tenfold cut in token-processing cost. The Groq 3 LPU, in liquid-cooled racks of 256 units, is slated for the second half of the year.
// KEY TAKEAWAYS
This week's throughline is AI moving from demo to consequence. A frontier model crossed OpenAI's own High cyber threshold and immediately proved it by finding real Chrome zero-days — capability gated behind hardware keys and vetted access, which is itself the story. Regulation stopped being theoretical the same week, with the EU AI Act's transparency and prohibition rules now carrying fines up to 7% of global turnover. Meanwhile the economics keep compressing: Meta put a genuinely capable 30B agentic model under Apache 2.0 on a single consumer GPU, Anthropic made cheap Sonnet 5 pricing permanent, and Nvidia is retooling silicon around agent runtime loops rather than training runs. Money is chasing the same shift — Cognition's jump toward $40 billion is a bet on coding agents, not chatbots.