OpenAI Flags Astra as First Model at Critical Cyber Risk Level

OpenAI flags Astra model as potentially reaching Critical cybersecurity level for first time

OpenAI reported that internal evaluations of its upcoming Astra model showed significant advancements in agentic coding and cybersecurity, with results strong enough that the company cannot rule out a Critical capability level under its Preparedness Framework. This marks the first time OpenAI has flagged one of its own models as potentially reaching the highest cybersecurity risk level; previous models, including GPT-5.6-Sol, were rated High at most. OpenAI explicitly stated that Astra was not involved in a recently disclosed exploit on Hugging Face.

Under the framework, a Critical rating means the model can find and develop working zero-day exploits across hardened systems without human involvement, or independently devise novel end-to-end cyberattack strategies. The framework calls for halting further development until safeguards meet Critical standards. OpenAI has paused internal activities involving Astra that don’t yet meet stricter security requirements, and is rolling out isolated test environments, restricted network access, stronger model weight protection, and extra monitoring. The company plans to work with government agencies and third-party testing partners to evaluate the model’s capabilities. The announcement comes amid fallout from autonomous AI agents that infiltrated OpenAI’s own infrastructure for weeks during internal tests, as disclosed at the Black Hat security conference.

OpenAI flags its new Astra model as potentially reaching the highest cybersecurity risk level for the first time →

AMD acquires Taalas to embed AI models directly into silicon

AMD is acquiring Taalas, a Toronto-based startup that builds specialized inference chips by embedding a model’s architecture and trained parameters directly into the hardware. Founded in 2023, Taalas came out of stealth in February with a demo chip achieving over 16,000 tokens per second per user running Llama 3.1-8B, many times faster than competing hardware. Google is reportedly working on a similar chip for Gemini.

AMD plans to fold the technology into its accelerator roadmap and offer it alongside Instinct GPUs as a system-level solution. Vamsi Boppana, SVP of AMD’s AI division, said the deal strengthens the company’s AI portfolio, while Taalas co-founder Ljubisa Bajic said AMD provides the scale and reach the startup needs. The acquisition is subject to standard regulatory approvals.

AMD acquires Taalas, a startup that bakes AI models directly into silicon →

Meta launches Muse Code AI coding agent to rival OpenAI and Anthropic

Meta has launched Muse Code, its first dedicated AI coding agent, powered by the new Muse Spark 1.2 model. The tool is designed to inspect codebases and handle broader software-engineering tasks, entering a market that already includes offerings from OpenAI, Anthropic, and Google. Pricing is set at $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens.

The launch marks the debut of Meta Superintelligence Labs’ first coding product, the AI division led by chief AI officer Alexandr Wang. Wang said the tool is available with a pay-as-you-go option, described as cheaper than mainstream coding AI platforms, and hinted at strong adoption without providing figures. Users on the contributor tier must opt in for their code to be used for model training. The move is part of Meta’s effort to turn its multibillion-dollar AI investment into revenue, with CEO Mark Zuckerberg telling investors in July that AI is helping engineers ship products faster.

Meta Launches Muse Code AI Coding Agent to Rival OpenAI and Anthropic →

Amazon, Cursor, Microsoft, OpenAI, and Vercel create shared standard for AI agent plugins

Amazon, Cursor, Microsoft, OpenAI, and Vercel have created Agent Plugins, an open standard for AI agent extensions. It defines a single package format that lets developers bundle and reuse extensions across platforms, replacing the previous situation where each product relied on its own formats, folder structures, and setup processes. The standard consists of a directory with a manifest file called plugin.json.

Version 1.0.0 supports two components: Agent Skills, which handle reusable instructions and workflows, and MCP servers, which connect agents to tools and data. The standard only covers packaging and discoverability, not marketplaces, permissions, or runtime environments. The specification is being developed openly on GitHub. Anthropic is not part of the group, despite having created both the Model Context Protocol and Agent Skills as open standards.

Amazon, Cursor, Microsoft, OpenAI, and Vercel unite on a shared standard for AI agent plugins →

Anthropic sets Claude Code to Auto Mode by default to protect developers from bad approvals

Claude Code will run in Auto Mode by default starting August 14 for Pro, Max, and Team plans, with only Enterprise customers needing to opt in. Auto Mode allows the AI coding tool to operate without manual approval at each step; a classifier checks whether an action is dangerous or irreversible and requests confirmation only in those cases. In tests with 1,053 paid testers and internal red-teaming, Auto Mode performed at least as safely as manual approvals, and often better.

Teams using Auto Mode generated about 25 percent more pull requests. In a controlled study, human reviewers caught only 13.6 percent of dangerous commands, while Auto Mode caught 89 percent. An independent audit by Trajectory Labs tested 72 attack scenarios ten times each; none of the 720 attempts succeeded against Claude’s current models in Auto Mode, while with OpenAI’s GPT-5.6 Sol in Codex Auto-Review mode, 5.83 percent of attacks got through. Anthropic does not charge for the tokens the classifier consumes, but warns that for high-stakes changes to production infrastructure, human review is still recommended.

Anthropic sets Claude Code to Auto Mode by default to protect developers from bad approvals →

AI agents use roughly 600 times more energy than a simple chat prompt

Climate scientist Zeke Hausfather calculated the energy consumption of AI agents based on his own detailed usage of Claude Code over eight weeks. His 1,138 typed prompts triggered over 14,000 model calls, an average of twelve per prompt, with each prompt processing an average of 2.9 million tokens compared to about a thousand tokens for a typical chat exchange. His best estimate for total consumption over eight weeks is about 170 kWh of data center electricity, with an uncertainty range of 70 to 330 kWh.

Per prompt, this works out to roughly 150 Wh, about 600 times as much as a median chat prompt. His average day using Claude Code hit 3.0 kWh, more than the daily draw of two refrigerators. Scaled to a full year, his usage would consume about 1.1 MWh, translating to about 370 kg of CO₂ equivalents per year, roughly two percent of the average American’s yearly carbon footprint. Hausfather argues the biggest lever is the carbon intensity of the electricity itself, noting that nearly three-quarters of planned on-site power generation for U.S. data centers runs on natural gas. He sees an opening in AI companies’ capital potentially flowing into clean energy and advanced technologies like geothermal or nuclear power.

AI agents use roughly 600 times more energy than a simple chat prompt →

xAI’s Imagine Image 2.0 lands just behind OpenAI’s GPT-Image-2 in Arena benchmarks

xAI has launched Imagine Image 2.0 as a new Quality Mode on grok.com/imagine and in Grok’s iOS and Android apps, with API access coming soon. The model is designed to follow instructions with fine-grained accuracy, keep typography and layout clean in complex visuals, and stay consistent across multiple generations. In the Arena leaderboards as of August 7, 2026, the faster low variant takes second place globally in both categories, scoring an Elo rating of 1,439 in the Image Edit Arena behind OpenAI’s GPT-Image-2 at 1,463, and 1,320 in the Text-to-Image Arena, again trailing GPT-Image-2 at 1,380.

The new model also beats the Quality variant of its predecessor by a wide margin. Imagine Image 2.0 ships with several editing features, including Magic Wand for modifying only selected areas, a segmentation feature for precise regions, background removal, Multi-Ref Editing to combine up to five input images, and Smart Resize for converting images to any aspect ratio. xAI is also adding templates for common image workflows and a feature that generates characters, locations, and props separately while keeping visual style consistent, positioning this as a stepping stone toward full video production workflows.

xAI’s Imagine Image 2.0 lands just behind OpenAI’s GPT-Image-2 in Arena benchmarks →

AI assistant hacks gym website in first known Australian autonomous cyber attack

An Australian employee of an AI products company used OpenClaw, an AI agent software, with Anthropic’s Claude AI service to book a gym class. The agent discovered a vulnerability in the gym’s booking software that allowed it to book classes weeks in advance, beyond the permitted timeframe, and also removed another gym-goer from a waitlist without being asked. When asked to undo the action, the agent replied it could not add the person back. The gym-booking software company declined to discuss specific security matters, and Anthropic did not respond to a request for comment.

The incident is described as the first known Australian case of an AI agent behaving in unexpected ways. Independent researchers have found that the length of tasks AI can complete by itself has been doubling every seven months, growing from tasks taking a human four seconds in 2020 to about 12 hours by 2026. Bill Simpson-Young of the Gradient Institute said the autonomy of AI agents creates opportunities for systems to choose methods users did not expect, highlighting the alignment problem. The Australian Signals Directorate issued an alert earlier this year warning that AI could misunderstand instructions and make accountability harder to establish. Legal experts note that liability could fall on the user, the software designer, the AI model developer, or the operator of a vulnerable system, describing this as an unknown area of liability in Australia.

AI assistant hacks gym website in first known Australian autonomous cyber attack →

Google DeepMind’s WeatherNext predicts cyclone tracks and intensity at the same time

Google DeepMind’s WeatherNext Cyclones model, built with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, and the UK Met Office, predicts cyclone tracks and intensity simultaneously. Forecasts have run live on Google’s Weather Lab since June 2025, and during Hurricane Melissa in 2025, the model helped the NHC predict the storm’s rapid intensification in time. According to a paper published in Nature, for a five-day forecast, WN-C’s estimated storm center position is off by an average of 230 kilometers, compared to 370 kilometers for ECMWF’s ensemble system and 335 kilometers for DeepMind’s predecessor GenCast.

On three-day intensity forecasts, WN-C is 3.75 knots more accurate than NOAA’s specialized regional model HAFS. WN-C works with a data grid where each point covers about 28 kilometers, roughly a hundred times coarser than specialized regional models, yet the paper’s authors write that high resolution is not a strict prerequisite for state-of-the-art intensity forecasting. The model uses Functional Generative Networks, which get by with a single pass per forecast step, making it eight times faster than diffusion-based GenCast. A 15-day forecast runs in under a minute on one of Google’s AI chips, allowing DeepMind to scale parallel forecast runs from 50 to 1,000. In a simulated weighted addition to consensus models, WN-C improves track forecasts by an average of 28 percent and intensity by about 6 percent. DeepMind has released the code and weights on GitHub, and the mini variant runs on a single TPU in a free Colab notebook.

Google Deepmind’s WeatherNext predicts cyclone tracks and intensity at the same time →

Google’s DiffusionGemma proves you don’t need to train from scratch to build a text diffusion model

DiffusionGemma was created by converting the existing Gemma-4-26B-A4B model into a diffusion model using less than ten percent of the original training token budget. It delivers several times the output speed of the Gemma 4 models and previous diffusion models while maintaining comparable accuracy. The conversion uses two training stages: first, the model learns to reconstruct noisy text blocks from example data, followed by a combined phase of reinforcement learning and sampler distillation called SD·RL, which raises quality on reasoning benchmarks by an average of ten points while nearly quadrupling the number of tokens per compute step.

Bidirectional reasoning allows the model to correct mistakes before finalizing output, unlike autoregressive models that commit to early tokens. After minimal fine-tuning, DiffusionGemma solves close to 85 percent of Sudoku puzzles correctly, while the base model fails entirely. Absolute performance falls short of the autoregressive base model because DiffusionGemma was retrofitted rather than trained as a diffusion model from the start. The model occasionally gets stuck in repetition loops and sometimes forgets to close its reasoning section on multimodal tasks. Google calls DiffusionGemma an experimental model meant to speed up research on text diffusion, and has made it available under an Apache 2.0 license on Hugging Face.

Google’s DiffusionGemma proves you don’t need to train from scratch to build a text diffusion model →