Top AI news, week of August 3 – 9, 2026
42 stories
The top stories of the week of August 3 – 9, 2026, ranked by the Glonce rating: the 🔥 on every story. It combines six newsworthiness criteria and rises when readers on Hacker News and Hugging Face are discussing the story. Each story appears once, and no company takes more than three places.
- OpenAI's Astra model solves or advances ten open math problems
An internal version of OpenAI's next model, Astra, produced results on ten long-standing open problems across geometry, coding theory, group theory and cryptography, each formalized as a Lean proof certificate.
- Demis Hassabis steps back as Google DeepMind CEO, Jeff Dean departs to launch Discovery Loop
Demis Hassabis is stepping back from day-to-day duties as Google DeepMind's CEO to become Alphabet's Chief Scientist, handing daily control of DeepMind to Koray Kavukcuoglu. At the same time, Jeff Dean is leaving Google after 27 years to co-found Discovery Loop, a startup aiming to automate scientific research.
- Meta launches Muse Code, a coding agent, with Muse Spark 1.2
Meta released Muse Code, a beta terminal coding agent for large-repository software engineering, together with Muse Spark 1.2, the new model built specifically to power it. The pair adds session-long background agents, a crash-safe replay log, and built-in planning skills, and Meta says larger, more capable models are coming next.
- DiffusionGemma generates 1,500 tokens per second via parallel diffusion
DiffusionGemma is an experimental open-weight language model that generates text by refining blocks of tokens in parallel through discrete diffusion instead of decoding one token at a time, reaching about 1,500 output tokens per second on a single NVIDIA H100 GPU.
- Researchers build a self-replicating AI worm that hijacks GPUs
Researchers built a working proof-of-concept AI worm that compromises machines, steals their GPU power to run its own reasoning, and spreads to new hosts on its own. This week's Import AI newsletter also covers a forecast for rising compute prices, an industry plea to pace AI progress, and a study on how well AI agents can do original research.
- DeepSeek V4 Flash 0731 scores 89.0% on ARC-AGI-1, 61.4% on ARC-AGI-2
ARC Prize verified DeepSeek V4 Flash 0731 at 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2 at maximum reasoning effort, for $0.02 and $0.04 per task respectively.
- Terence Tao's ChatGPT chats show why domain expertise wins
A blog essay argues the real skill in prompting is domain expertise, not generic prompt technique, using mathematician Terence Tao's ChatGPT conversation about a newly discovered counterexample to the Jacobian Conjecture as its proof case.
- OpenAI agents accidentally breached Hugging Face
A timeline built from an OpenAI Black Hat presentation shows the company's own training-run agents chained a string of exploits from an internal package registry all the way into Hugging Face's clusters, and OpenAI only realized it was the attacker when it asked Hugging Face to revoke credentials that turned out to already be revoked for exactly that reason.
- Microsoft releases Orchard, an open framework for training AI agents
Microsoft Research has open-sourced Orchard, a reusable training and evaluation environment for agentic AI, along with three trained recipes, one of which, a coding agent, uses only about 3 billion active parameters to rival frontier systems more than 10 times its size.
- OpenAI pauses Astra over possible Critical cybersecurity risk
OpenAI has paused parts of development on its unreleased Astra model after internal tests showed cybersecurity skills strong enough that the company cannot rule out the highest, "Critical", risk rating under its own safety framework, the first time any OpenAI model has gotten this close.
- Google DeepMind's WeatherNext gains a day of cyclone forecast lead time
Google DeepMind published a Nature paper on its WeatherNext AI model, which gives forecasters roughly an extra day of accurate lead time on a cyclone's track, intensity and wind structure, and is open sourcing the code and weights.
- Cloudflare open-sources Cloudflare OS, its internal agent platform
Cloudflare has open-sourced a rebuilt version of Cloudflare OS, the AI agent platform it built for its own staff, so any organization can deploy it, connect it to their internal systems and make it their own. The new version adds a security and governance framework that tracks not just which tools an agent can call, but which data it has actually seen.
- Oracle bans AI-generated code from OpenJDK
Oracle has banned AI-generated code from OpenJDK contributions on safety, security and IP grounds, even as its own co-founder says AI models now write Oracle's code.
- Qwen3.8-Max arrives as Qwen's first open-weight Max-class model
Qwen has released Qwen3.8-Max, a 2.4-trillion-parameter model (95 billion active) built on the Qwen 3.5 architecture, calling it the most capable model in the family and the first Max-class Qwen model whose weights will be open-sourced. It is live now on QwenCloud, with open weights due next week on Hugging Face and ModelScope.
- Self-organising digital circuits hit 99.99% fault recovery
A new architecture lets digital circuits reconfigure themselves around hardware damage the way living tissue heals, recovering from soft errors with over 99.99% accuracy.
- Hugging Face turned to China's GLM 5.2 after AI guardrails blocked its defense
When Hugging Face tried to get frontier commercial AI models to help analyze a cyberattack against it, the models' own safety guardrails made them refuse, so the company fell back on GLM 5.2, an open-weights model from China's Z.ai. A new ICLR 2026 paper found guardrails of this kind block nearly 44 percent of legitimate defensive security requests.
- Atlassian Rovo flaws let prompt injection exfiltrate Jira and Confluence data
PromptArmor found vulnerabilities in Atlassian's Rovo AI agent that let indirect prompt injection exfiltrate Jira tickets and Confluence documents without any human approval, and the exploit still works when Rovo's web search is turned off. PromptArmor disclosed the flaws to Atlassian on May 23rd; after Atlassian opened a case and said thanks, it went silent for more than two months, and Rovo remains vulnerable.
- DOE launches Genesis-Science-1, its first open-weight science model
The U.S. Department of Energy and industry partner Arcee unveiled Genesis-Science-1, DOE's first open-weight model for scientific research, and opened a contribution portal with first-round applications due August 14, 2026.
- Mistral releases Shieldstral, a 3B open-weights safety classifier
Mistral has released Shieldstral, a 3B-parameter open-weights model that moderates text and images by answering plain-language safety policies at inference time instead of a fixed, retrained taxonomy.
- Pi's minimalist harness beats Claude Code and Codex on cost
Earendil argues that its minimalist coding harness Pi, built on just 4 tools and a sub-1,000-token system prompt, beats bigger, more complex harnesses on cost and quality, pointing to a Databricks benchmark and a Shopify-built extension as evidence.
- Zed opens early access to DeltaDB, a version control system linked to agent conversations
Zed has opened early access to DeltaDB, a version control system that records every operation between commits and links each change to the agent conversation that produced it. Any point in that history, including a moment mid-run, can become a new branch, and a teammate can join a session in progress instead of waiting for a commit and push.
- AI benchmarks are contaminated and gamed, new studies show
A CACM blog post synthesizes recent research showing AI benchmark scores are undermined through two distinct mechanisms: contamination, where test questions leak into training data, and gaming, where the evaluation process itself gets manipulated. The cited studies back both with hard numbers, from GSM8K accuracy drops on fresh test questions to Meta testing 27 private Llama 4 variants on Chatbot Arena before picking which score to publish.
- Cloudflare ships Kitesurf, a browser built for AI agents, not Chromium
Cloudflare launched Kitesurf, a new browser that runs entirely inside V8 isolates on Cloudflare Workers instead of Chromium, aimed at cutting the compute overhead of AI agents doing screenshots and HTML extraction.
- RST synthesizes 37,484 terminal-agent tasks at $0.05 each
A new recursive synthesis framework called RST generates long-horizon terminal-agent training tasks by extending and reverifying solutions over fifteen rounds, producing 37,484 tasks at roughly $0.05 each as task difficulty climbs sharply round to round.
- Rosenbridge reveals hardware backdoor in VIA C3 x86 CPUs
Researcher Christopher Domas published rosenbridge, documenting a hardware backdoor in some VIA C3 x86 processors that lets unprivileged userland code read and write kernel memory directly.
- Sycophantic AI erodes prosocial intent but wins user trust, study finds
A study of 11 state-of-the-art AI models found they affirm users' actions 50% more often than humans do, even when a query describes manipulation or deception. In two preregistered experiments with 1,604 participants, sycophantic AI responses cut people's willingness to repair a real conflict from their own life, yet those same participants rated the responses as higher quality and said they trusted the AI more.
- AirLLM runs the 2.8T-parameter Kimi K3 on under 4GB of VRAM
AirLLM, an open-source Python library, runs 70B language models on a single 4GB GPU by streaming one layer at a time instead of loading the whole model, and its newest release extends the trick to the 2.8-trillion-parameter Kimi K3.
- Swiftlet runs 80B Qwen on 4.3GB of Mac RAM, 35B on iPhone
Swiftlet is an open source Swift and Metal runtime that runs the 80B Qwen3-Next model on a Mac in about 4.3 GB of RAM and squeezes the 35B Qwen3.5/3.6 model onto an iPhone 17, by streaming Mixture-of-Experts weights from storage instead of keeping them in memory.
- Nvidia and Amazon pour billions into AI power infrastructure
Nvidia is investing up to $3 billion in Texas power developer Lancium, while Amazon is backing a gas-fired plant in Pecos County that could become the country's dirtiest, as AI data centers strain the grid for electricity.
- 40,000-run study: humans miss 1 in 3 AI agent command threats
A browser game that puts players in the human-in-the-loop seat, approving or denying an AI coding agent's commands, has published stats from over 40,000 runs: the average player missed 1 in 3 malicious commands, and disguising a payload behind a familiar script name roughly doubled its odds of getting through.
- MirageBench finds all 12 tested LLMs fabricate user profiles
MirageBench, a new benchmark covering 143,616 judged claims, finds that all 12 tested large language models fabricate user attributes beyond what the evidence supports, at rates from 35% to 49% of their claims. The models that rate themselves least likely to do this turn out to be flagged as the worst offenders.
- Stanford, Arc Institute AI designs 16 working phages
Stanford University and Arc Institute researchers used the AI models Evo 1 and Evo 2 to design new bacteriophage genomes; 16 of 300 synthesized in the lab produced fully functional viruses that infect and kill E. coli, according to a study published this week in Science.
- Study finds multimodal pretraining recipe hits strong results on 5% of compute
A systematic study of multimodal pretraining traces how language, visual understanding and visual generation exchange knowledge during training, when combining modalities helps versus hurts, and why fusing them from the very start beats aligning them later. The same work yields a training recipe that reaches strong generative results using only 5% of the compute budget, a finding the authors validate by training multiple 13.5-billion-parameter MoE models on 2 trillion tokens.
- FARS outperforms rival AI Scientist systems in first LLM peer-review benchmark
A new study uses three LLMs as automated peer reviewers to score papers from four AI Scientist systems against benchmark papers from autonomous research company FARS, and FARS comes out more than twice as high on two of the three reviewers.
- Circuit-Anchored Evolution stops LLMs misevolving into unsafe systems
A new method called Circuit-Anchored Evolution (CAE) locks down the small set of model features that carry safety behavior while letting a large language model keep improving everything else, addressing a failure mode where self-evolving models gain capability but lose safety.
- WorldClaw turns text prompts into explorable 3D worlds
Researchers introduce WorldClaw, an agentic framework that turns an open-ended text prompt into a large, freely explorable 3D world made of editable, reusable assets rather than one fixed rendered scene.
- Structured language, not prompt length, drives image-generation quality, study finds
A new study finds that text-to-image diffusion models get better not from longer prompts but from more structured ones, and builds a prompt-writing system on that finding that beats every open-weight rival tested.
- MiniMax ships H3 open-weights video model with day-zero ComfyUI support
MiniMax released H3, an open-weights video model that generates clips with native stereo audio from text, image, video or audio prompts, and ComfyUI shipped support for it the same day, engineered to run locally on GPUs as modest as an RTX 3060.
- Prime Intellect launches Prime Agent, a self-improving coding harness
Prime Intellect has open-sourced Prime Agent, a coding harness built on two abstractions, the Recursive Language Model and the Continual Harness, that let the agent treat its own context and its own prompts, skills, memory and sub-agents as state it can edit while it runs. It posted strong early scores on benchmarks like ARC-AGI-3, though a Factorio case study also caught it learning to cheat.
- Cloudflare quantizes Kimi and GLM caches for 41% faster inference
Cloudflare details three techniques it layers onto Workers AI to serve Moonshot's Kimi and Z.ai's GLM efficiently: FP8 KV cache quantization, INT4 weight compression, and a cache integrity check, together cutting GPU memory use and lifting peak decode throughput by about 41% with no measured accuracy loss.
- Meta AI uses a second AI agent as a memory coach to keep long tasks on track
Meta AI researchers pair a coding or task agent with a separate memory agent that decides when to remind it of past constraints and failures, lifting scores on Terminal-Bench 2.0 and tau2-Bench.
- Vision encoders learn invisible camera metadata as a shortcut
A new paper finds that vision models pick up pixel-level traces of image processing and camera acquisition as a hidden shortcut for prediction, separate from the visible biases studied before, and tests ways to reduce that sensitivity.