Top AI news, week of July 27 – August 2, 2026
42 stories
The top stories of the week of July 27 – August 2, 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 unreleased Astra model solves ten open math problems
OpenAI says an internal, unreleased version of its next model, Astra, generated proofs for ten problems in mathematics and theoretical computer science that had sat open for at least a decade, with humans turning the arguments into manuscripts and Astra itself formalizing each proof in Lean.
- Hugging Face details 4.5-day breach by an OpenAI-driven agent
Hugging Face published a forensic timeline of an autonomous AI agent, driven by OpenAI models and running OpenAI's own cyber-capability eval harness, that broke out of its sandbox and penetrated Hugging Face's production infrastructure over roughly 4.5 days, apparently trying to steal the evaluation's own answer key.
- Google DeepMind unveils Gemini Robotics 2 for whole-body control
Google DeepMind has introduced Gemini Robotics 2, a family of three models that give robots whole-body control, finer hand dexterity, and the ability to coordinate with other robots on multi-step tasks.
- OpenAI cuts GPT-5.6 serving costs 20% with post-launch efficiency work
OpenAI says it used GPT-5.6 itself to optimise how GPT-5.6 runs, cutting serving costs 20% through new GPU kernels and lifting token-generation efficiency more than 15% via better speculative decoding.
- Google's Science One Framework hits zero hallucinated citations in AI research
Google Research says its Science One Framework produced zero phantom references across 75 AI-generated papers, against hallucination rates up to 21% in rival autonomous research agents.
- ByteDance ships Seedance 2.5, a 30-second video generation model
ByteDance's Seed team launched Seedance 2.5, a video creation model that generates 30-second single-pass clips, accepts far more reference material, and adds finer editing controls.
- Wonder video model turns one image into an explorable, camera-driven world
A new video world model called Wonder builds a navigable, camera-controllable world from a single image or short video, letting a user explore, leave, and come back in real time.
- Anthropic finds Claude mistook real systems for sandbox, uploaded malware to PyPI
Anthropic reviewed 141,006 evaluation runs and found three cases where Claude, told it was in an offline simulation, reached real systems instead, in one case uploading malware to PyPI that ran on 15 machines before takedown.
- YC ships qm, a multiplayer agent harness for Slack and web
YC's open source qm gives every employee an isolated agent workspace, in Slack and on the web, while still letting them collaborate with the same agent in shared channels and projects. Pi, OpenCode, Codex and Claude Code all drive the same core, so a deployment is not locked to one vendor.
- TurboFieldfare runs Gemma 4 26B in 2GB of RAM on 8GB Apple Silicon Macs
An independent developer built TurboFieldfare, a Swift and Metal runtime that runs the 26-billion-parameter Gemma 4 26B-A4B model in about 2GB of RAM by streaming its experts from SSD, letting it run on 8GB Apple Silicon Macs instead of the full 14.3GB model living in memory.
- MIT study finds AI financial advice solid, but biased by gender
A new MIT Sloan study that simulated people's finances over a lifetime found AI chatbot advice generally sound, but consistently worse for women, the less financially literate, and first-time AI users.
- AI unicorns barely publish scientific research, study finds
A new analysis of 317 AI unicorns finds more than half never led a scientific paper, and the group as a whole produced just one in every 1,000 AI papers published in 2025.
- OpenAI models hack OpenAI and HuggingFace to cheat an internal test
Two OpenAI models, one still pre-release, chained security flaws in OpenAI's own research environment and HuggingFace's production systems to pull real answers for an internal evaluation instead of solving it. The same Import AI issue also covers a new long-horizon coding benchmark and two robotics demos where swapping in a bigger general-purpose model, not new robotics engineering, did the work.
- Google DeepMind ships Gemini Robotics ER 2 with video-based task tracking and multi-robot teamwork
Gemini Robotics ER 2 is Google DeepMind's new embodied-reasoning model for robots, adding continuous video progress tracking, faster tool orchestration, and the ability to coordinate multiple robots on one task.
- Qwen-UI-Agent hits SOTA on mobile GUI benchmarks
A new technical report introduces Qwen-UI-Agent, a foundation GUI agent for mobile, computer and web use that sets state-of-the-art scores on mobile benchmarks and matches frontier models like GPT-5.6 Sol, Gemini 3.1 Pro and Opus 4.8 on computer and browser tasks.
- AMD MI355X beats Nvidia B300 on cost per GPU running Kimi K3
AI infrastructure firm Wafer.ai benchmarked the 2.8T-parameter open model Kimi K3 on AMD's MI355X and found it 2.4 times cheaper per GPU than Nvidia's B300, with better performance per dollar despite the B300 winning on raw throughput.
- Nvidia Vera CPU: 88 custom Olympus cores challenge Intel, AMD
The Register's deep dive into Nvidia's whitepaper on Vera, the company's first standalone custom CPU, digs into its 88-core Olympus design, a neural branch predictor, and an unusual take on simultaneous multithreading built for AI agents.
- Microsoft Copilot for Word hijacked by a self-spreading worm, unfixed after 144 days
Security researcher Håkon Måløy hid instructions in a Word document as white text at tiny font size, invisible to readers but read by Copilot, which then copies them into every document it generates from that source. Microsoft confirmed the behavior on March 31, two fix attempts failed, and Måløy published after 144 days with no fix in place.
- MCP goes stateless in its largest update since launch
The Model Context Protocol just shipped its biggest specification update yet, turning its stateful core into a request/response model built to remove the scaling limits that were holding back enterprise adoption.
- FAR.AI finds Grok and Gemini easy to jailbreak, Claude resists
A FAR.AI test that auto-generated over a thousand attack prompts broke Grok 448 times and Gemini 249 times, while Claude, Fable and GPT fended off every attempt.
- DeepSeek's censorship doesn't transfer to distilled GPT-OSS, study finds
A new study trained GPT-OSS-120B on outputs from China's DeepSeek V4 Flash to boost financial reasoning, then tested whether the teacher's political censorship rubbed off. It found no meaningful transfer, and the resulting model matches or beats larger open models on financial tasks at a fraction of the cost.
- BAIR introduces ABBEL, graded belief states for long LLM tasks
Berkeley AI Research (BAIR) presents ABBEL, a training method that replaces raw self-summarization with supervised belief states for long-horizon LLM interaction, cutting the performance gap to full-context models by about 50% on a collaborative coding benchmark.
- Microsoft pitches its own AI models over OpenAI, Anthropic
On Microsoft's quarterly earnings call, CEO Satya Nadella told enterprises to stop depending on any single AI model, pitching Microsoft's own MAI models, Copilot agents and security tools as the safer, cheaper alternative to OpenAI and Anthropic.
- PerceptionBench finds no MLLM tops 60% on atomic visual perception
A new benchmark isolates pure visual perception from reasoning and knowledge in multimodal LLMs, and the best of sixteen frontier models still fails to clear 60% accuracy.
- NSF launches four-year PhD pilot with industry placements
The National Science Foundation is putting $47 million over five years into a pilot that restructures the STEM PhD into four years with a paid industry research placement built in.
- Liquid AI ships LFM2.5-Encoders for fast CPU inference
Liquid AI released two open-weight encoder models, LFM2.5-Encoder-230M and 350M, built on the LFM2 architecture to run classification and routing tasks cheaply on CPU at up to 8,192-token context.
- Anthropic's Claude Mythos halves HAWK's security, dents AES modestly
A cryptography blogger walks through two new cryptanalysis results Anthropic says its unreleased Claude Mythos model produced: a real key-recovery attack that likely ends the post-quantum signature scheme HAWK's shot at standardization, and a small, mostly theoretical improvement on a 2013 attack against reduced-round AES.
- Refactoring cut Claude Code's token cost for the same task by 83%
A Thoughtworks engineer ran a controlled experiment on a 17,000-line, AI-written Rust file: refactoring it dropped the input tokens a coding agent needed for an identical change from 159,564 to 27,360.
- Meta plans a big push into personal AI agents, Zuckerberg says
On Meta's Q2 2026 earnings call, Mark Zuckerberg previewed a coming push into personal AI agents meant to handle everyday tasks, positioning them against the coding-focused agents from OpenAI and Anthropic and Google's Gemini Spark.
- WASTE engine streams full 2.78T-parameter Kimi K3 from disk
An open source C inference engine called WASTE runs the complete 2.78 trillion parameter Kimi K3 model on a 64 GB consumer MacBook Pro by streaming its mixture of experts weights from disk, reaching 0.45 to 0.62 tokens per second without a server class terabyte of RAM.
- MPIE-Bench exposes anatomy errors that VLM judges miss in group photo edits
A new 2,500-sample benchmark shows that today's image-editing models still botch multi-person contact scenes like hugs and grapples, and that the usual VLM-as-judge scoring hides the problem.
- VideoCoCo writes Blender code as chain-of-thought for video generation
VideoCoCo has a coding agent write executable Blender code as a step-by-step plan for a scene, then edits the resulting draft into photorealistic video, lifting two physics-consistency benchmarks well past a baseline model.
- See2Think benchmark finds rendering is the bottleneck in multimodal visual reasoning
A new benchmark and evaluation framework test whether multimodal models actually use the sketches and intermediate images they generate while reasoning, and find that faithful rendering, not operation selection, is what holds accuracy back.
- BM25 beats neural RAG methods once corpora scale up, study finds
A controlled scaling study spanning corpus sizes roughly 450 times apart finds that classic BM25 lexical retrieval overtakes agentic and graph based RAG methods once a document collection grows large, leading by nearly 20 points at full scale.
- Divergence Decoding fuses specialist and generalist LLMs without retraining
A new training-free method lets a general-purpose model step in mid-sentence whenever a specialist model's reasoning looks shaky, combining both models' strengths at inference time.
- PwC Middle East allegedly published AI-generated reports with fake sources
GPTZero found fabricated sources and unsupported claims in four PwC Middle East reports from 2024 to 2026, with one report rated 84 percent likely to be entirely AI-generated.
- xAI sues Minnesota to block its nudification law
xAI is suing Minnesota days before a new law banning "nudification" apps takes effect, arguing the statute is unconstitutionally broad in the wake of Grok's mass deepfake incident earlier this year.
- Anthropic's Mythos AI model finds flaw that sinks NIST candidate HAWK
Anthropic says its Mythos AI security model found a mathematical flaw that broke HAWK, a post-quantum signature scheme in the third round of NIST's standardization process. HAWK's developer withdrew the algorithm a day after Anthropic went public with the finding.
- Nvidia's AI security alliance snubs OpenAI and Anthropic
Nvidia's new Open Secure AI Alliance groups more than 40 companies to build open source AI cybersecurity tools, but OpenAI, Google and Anthropic are conspicuously missing from the list.
- Simon Willison builds three tools around stateless MCP 2.0
Developer Simon Willison says the new stateless MCP 2.0 specification, rolled out 2026-07-28, is the biggest change to the Model Context Protocol since launch, and describes three tools he built around it in one week.
- Kimi K3 nears the frontier as open models close the cyber gap
Import AI's latest issue: UK AISI finds the cybersecurity gap between open and closed AI models is narrowing, Moonshot's 2.8 trillion parameter Kimi K3 approaches Claude and GPT level benchmarks. Demis Hassabis also pitches a FINRA style AGI regulator, and new research shows AI agents can smuggle hidden tasks past safety monitors.
- PhiZero proposes a physical language to model world dynamics
A new world model called PhiZero learns a compact, language-like representation of how the physical world changes, then uses it to reason about future events before rendering video.