Frontier
Papers, benchmarks, and technical breakthroughs with real downstream impact.
Sam Altman says OpenAI still plans broad access, but the lab is tightening controls around a model with potentially critical cyber capabilities.
Hidden inside Grok Build 1.0, the Computer Hub command can turn a developer's local workspace into a remotely accessible tool server.
The lab is tightening internal security around its next major model after saying it cannot rule out critical cyber capabilities.
Google's Gemini Robotics 2 controlled Apollo 2 in three configurations, with reported dexterity results ranging from 32% to 92%.
A GitHub issue opened by an account with no repository privileges was enough to execute code on the CI runners behind Anthropic's and Google's own coding-agent repositories.
The Toronto chipmaker raised $219 million to hardwire AI models into silicon; AMD did not disclose the price and plans to pair the technology with Instinct GPUs.
The release focuses on reliability and controls, while Grok 4.5, workflows and an open-source harness carry the bigger bet.
The August round includes follow-on investments from Coatue and Nvidia, new capital from Blackstone vehicles and participation from Jane Street, following a US$505 million equity investment announced in April.
The classifier rewrite targets everyday health and clinical queries, while virology, toxicology and molecular design still route to Opus 5.
The upload confirms RuntimeWire's earlier reporting from closed captions obtained through a QR code leaked from Greg Brockman's personal X account.
Lightspeed and Committed Capital co-led the Series A as Gravity tests ads that influence software agents before users ever see them.
The Sora 2 contributor is betting that a desktop interface can bring coding-agent workflows to the rest of knowledge work.
The founders say Atlas can simulate physical environments in real time while using a fraction of the compute required by larger video models.
Haz Hubble and Wyatt Lansford are running the YC-backed company as a two-founder team, with Pally valued at $30 million.
Free and Go users will get unlimited everyday chats and a Think button, while Plus and Pro users get an updated Sol model with adjustable reasoning.
The working draft packages Agent Skills and MCP servers so one plugin can run across competing AI clients, with Vercel also helping govern the specification.
Jeff Dean, a longtime senior Google executive, is departing to launch an AI research startup called Discovery Loop, joined by fellow Google veterans Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, with Dean reportedly set to serve as CEO.
The agreement turns APA from an outside reviewer into a partner on resources for families, clinicians and schools.
The account behind the claim calls Astra OpenAI's largest pretrain since GPT-4.5 and names "mewfour" as its release candidate.
The release includes WeatherNext 2 and WeatherNext Cyclones checkpoints, plus a lighter version for researchers and weather agencies with constrained compute.
The two-year-old startup says an unnamed hyperscaler is deploying its optical hardware under a multi-billion-dollar agreement.
Google will invest in and provide cloud services to Discovery Loop, while Koray Kavukcuoglu takes over Google DeepMind's daily operations from Demis Hassabis.
PortSwigger says HTTP Terminator, an artificial intelligence (AI)-assisted research system built by James Kettle, generated and proved new HTTP desynchronization techniques after exploring 30,000 candidate attack vectors.
The new Quality Mode adds region edits, five-image references and smart resizing, while API access remains pending.
Four Opus 4.6 agents beat one Opus 4.8 agent on a codebase benchmark, though a newer model has already edged past the result.
The free extension combines a story feed, model-release tracker, focus prompt, synced task list and rotating wallpapers.
One model generated the assignment; the other kept defaulting to an unrelated couch-bound talking-head clip. Across all four tests, MiniMax H3 Reference to Video didn’t just edge Heygen v5 Digital Twin — it exposed a category mismatch.
Economic Optimism, AI’s New Frontier, and Finding Your 25th Hour | DisrupTV Ep 447 Elle DeRosa Fri, 7 Aug 2026 - 15:25 R "Ray" Wang Principal Analyst and Founder Constellation Research R “Ray” Wang is the CEO of Silicon Valley-based Constellation Research Inc.
Chuck Feerick and Guru Nadiger are building an AI platform for payer workflows as CMS decision-time and electronic-interface requirements take effect.
ARC Prize's outside evaluation documents how Liang Wenfeng's open-weight model falls from 61.4% at Max effort to 46.0% at Low, with Max costing $0.04 per task.
The benchmark site's profiles drew users before Alex Cheema could announce them; a wider launch is set for August 13th.
Former Google executive Justin Inman is building software to monitor, predict and act on how brands appear in AI-generated answers.
Justin So says the YC-backed AI agent is live in more than 20 plants, where workers can report problems without installing an app.
One API and prepaid balance now cover hosted and third-party models; model-first failover and prompt-based routing remain in testing.
Research finds your AI generated security patch is more likely to fail than fully fix a vulnerability. It might even introduce brand new flaws to exploit along the way. The post More than half of AI-generated patches are broken appeared first on CyberScoop.
This matchup wasn’t especially close: DeepSeek-V4-Flash controlled the practical, instruction-heavy work, while cohere-command-a mostly stayed competitive only on a couple of writing tasks. The result is a decisive overall win driven by cleaner execution, stricter adherence to format, and fewer self-inflicted errors.
Airbnb graduates to AI native: A look at the metrics, ROI approach Larry Dignan Fri, 7 Aug 2026 - 03:45 Larry Dignan Editor in Chief of Constellation Insights Constellation Research Larry Dignan is Editor in Chief of Constellation Insights at Constellation Research, where he leads editorial coverage focused on...
Shield AI and Taiwan's National Chung-Shan Institute of Science and Technology (NCSIST) said on August 6, 2026, that they have successfully completed a live demonstration of a multi-drone cooperative autonomous search mission, using three Mighty Hornet III drones developed by NCSIST.
Kanjun Qiu and Josh Albrecht's AGPL tool beat a linear agent on nanochat, though its compute costs and research limits remain substantial.
The 124-billion-parameter mixture-of-experts model activates 5.1 billion parameters per token and supports native 256K context for tool-driven workflows.
William Tunstall-Pedoe says current AI can generate ideas but cannot reliably identify which ones are valuable, a distinction that underpins UnlikelyAI's neurosymbolic approach.
Frontier Security said the open-weight model accessed information outside an AI safety test environment, though the bypass method remains undisclosed.
Navya Tuteja, a rising senior at Thomas Jefferson High School for Science and Technology, created a free artificial intelligence tool that assists neurodivergent job seekers.
QwenCloud lists `qwen3.8-max` as a Token Plan-only preview whose capabilities may change, raising versioning questions for developers evaluating its visual and agent features.
Research from Amazon Ads indicates advertising platforms and regions which previously seemed out of reach to SMBs are now accessible — thanks to AI.
In a session attended by Ground Level AI, OpenAI researchers said the company is "consciously slowing down research to enhance security" while a full technical postmortem is still underway.
Google is restructuring the leadership of its AI division at a moment when the company can least afford instability at the top of its research organization, with chief scientist Jeff Dean departing after 27 years and Demis Hassabis stepping back from day-to-day control of Google DeepMind.
Two senior engineers are leaving company to launch startup amid fears Google is falling behind in AI race Sir Demis Hassabis is stepping down as chief executive of Google DeepMind, in a leadership overhaul of the UK-based AI research lab.
Apple GPU architecture alumnus Jacob Peake is targeting frontier labs with sealed chip-design tasks and physics-based grading.
The Drift co-founder will run Composer and Customer Agent after the Q3 deal closes, reuniting with Klaviyo co-founder Andrew Bialecki.
A new non-paper from the WE BUILD Consortium offers a policy vision for Europe’s chance to build a trusted identity layer for AI agents.
The August 7 session promises open H3 weights, stereo audio and ready-made workflows, but no official benchmark yet identifies the hardware, memory use or generation speed.
The milestone gives the independent AI and startup publication a growing base of readers who receive its reporting outside algorithmic feeds.
The agents rebuilt a deleted cross-run message board within 48 hours, restoring coordination before the July Hugging Face intrusion.
OpenAI resumed training two days after its models took over an internal server and caused an outage. Within 48 hours, an experimental model found another way for agents to communicate—and the activity later spilled into Hugging Fac
Mark Zuckerberg's terminal agent uses persistent subagents and Muse Spark 1.2 to plan, edit and validate code.
Dragonfly led the round six months after Sapiom's seed, as the startup expands from agent payments into routing, building and runtime tools.
OpenAI said it disrupted a Cambodia-based scam operation that used its generative artificial intelligence (AI) chatbot ChatGPT to facilitate a wide range of investment, romance, gambling, and law enforcement impersonation schemes.
SkillBench's July financing backs a privacy-first system for measuring AI coding gains, human contribution and the skills developers retain.
Ramin Hasani's latest LFM targets tool calling on phones, laptops and Raspberry Pi-class computers, with a license catch for larger businesses.
The former alignment researcher joined a San Francisco team training models on thousands of hours of non-invasive neural recordings.
The seed backs Colin Webb's model: sell warehouse labor now, collect training data on every shift, and automate the work over time.
Cybersecurity researchers have discovered more than half-a-dozen services advertisements for illegal access to artificial intelligence (AI) models on underground cybercrime forums and messaging platforms.
The Kaggle contest uses more than 5,000 exams from 16 institutions and includes a separate efficiency track.
The data.world co-founder used a July 24th panel with Jamie Metzl to argue that faster AI requires stronger human moral judgment.
Pieter Levels says a Meta employee described a plan to keep AI search queries from Google, adding an unverified motive to a project reported in 2024.
UK's AISI finds 19 instances where Anthropic's Mythos, OpenAI's GPT-5.6 Sol tried attacks Larry Dignan Wed, 5 Aug 2026 - 03:41 Larry Dignan Editor in Chief of Constellation Insights Constellation Research Larry Dignan is Editor in Chief of Constellation Insights at Constellation Research, where he leads editorial...
Cua Driver 0.19.0 binds exact tabs to native windows, letting agents cross web pages and desktop apps without taking the cursor.
Garry Tan's "personal AGI" thesis puts persistent memory and reusable skills, rather than model access, at the center of a founder's advantage.
Hosted-runner capacity constraints left CI jobs queued or failing for more than three hours during the US workday.
An agent running Anthropic's Claude Mythos 5 spent 34 hours trying to get a malware dropper merged into a real open-source project during a cyber evaluation by the UK's AI Security Institute.
The YC S26 team is betting computer control can automate hotel work without forcing operators to replace their existing software.
The Rust-based beta cuts CPU and memory use in Cloudflare's tests, while accepting slower loads and narrower web compatibility.
MiniMax's H3 video model is available through Luma's multi-model creative product, while pricing, eligibility, regional availability and developer access remain undocumented.
AISI recorded 19 out-of-scope actions during a July 28 cyber evaluation, including a Mythos 5 agent's use of fake identities to pressure an open-source maintainer.
Meta reported perfect physics theory scores, while leaving the models and evaluation setup unspecified.
The public beta doubles Wan's prior 15-second ceiling and starts at 5 cents per generated second through Alibaba Cloud and Qwen Cloud.
The multiyear contract is worth at least half Mirendil's seed financing, although the frontier AI lab has yet to release a public product, benchmark or customer deployment.
NVIDIA is opening Alpamayo 2 Super for commercial model development and redistribution, but has disclosed no customers or deployments.
Edo Liberty and Ash Ashutosh are expanding Pinecone beyond vector storage with a governed knowledge layer that runs in the customer's cloud.