GPT-6 Astra Ships as Nvidia Bets $13B on Hugging Face

AI & Machine Learning

OpenAI began rolling out GPT-6 Astra, the first model the company has designated “Critical” under its own Preparedness Framework after it scored 100% on the ExploitBench benchmark and showed it can autonomously discover unknown security flaws and build working exploits against well-defended systems without a human directing every step. The rollout is phased: a limited set of vetted organizations in OpenAI’s Daybreak cybersecurity program get access first, with broader availability following over the next several days across ChatGPT Plus, Pro, Business and Enterprise plans plus the API and AWS, priced at $10 per million input tokens and $50 per million output tokens. The public build refuses advanced offensive tasks such as generating proof-of-concept exploits, with those restrictions set to loosen only for vetted defenders through Daybreak. Enterprise administrators must manually enable Astra for their workspace, since access is off by default. Source: CSO Online Verified: True

Meta released Muse Spark 1.3, a model the company says closes most of the gap with the frontier labs on agentic and coding tasks, with Artificial Analysis scoring it behind only Claude Fable 5.1 and Claude Opus 5 on its Intelligence Index. The model is built to sustain longer-horizon work by asking clarifying questions and tracking multiple workflows in a single thread, and for coding it needs roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 while working with a 1M-token context window. It is priced the same as its predecessor at $1.25 per million input tokens and $4.25 per million output tokens, and is available immediately through Meta’s developer API and rolling out to Instagram, Facebook and Meta AI. Source: Meta AI Research Verified: True

Consumer Hardware

Samsung unveiled two new Bespoke AI washing machines ahead of IFA 2026 in Berlin: a 10kg model rated 70% more energy-efficient than Europe’s minimum Class A threshold, consuming just 15.3 kWh per 100 cycles, and a 13kg large-capacity model that still fits a standard 600mm European cabinet using Samsung’s SpaceMax drum engineering. Both carry large touchscreens, SmartThings connectivity for remote monitoring, and an Auto Open Door function that vents moisture after a cycle finishes to cut mold and odor buildup. Samsung is showing the lineup through September 6 at Humboldt Carré as the IFA Creator Hub’s main sponsor. Source: SamMobile Verified: True

Acer used the run-up to IFA 2026 to unveil the Swift Blade 14, a 799-gram Windows 11 laptop it is billing as the lightest 14-inch notebook sold globally, built with carbon-fiber outer covers and a magnesium-aluminum keyboard deck to hit that weight without sacrificing rigidity. The machine pairs a 3K OLED display with 100% DCI-P3 coverage and 4.3mm bezels with Intel’s new Wildcat Lake chips, ranging from a Core 3 304 up to a Core 7 350, alongside 12GB of LPDDR5 RAM. Acer has set EMEA availability for December 2026 at under €2,000, with a North American launch to follow in the first quarter of 2027. Source: Tom’s Hardware Verified: True

Cybersecurity

CISA added seven vulnerabilities to its Known Exploited Vulnerabilities catalog based on evidence of active exploitation, spanning Sangoma Switchvox (SQL injection), Kludex Starlette (HTTP request/response smuggling), Kestra OSS (OS command injection), BerriAI LiteLLM (improper authentication), JFrog Artifactory (improper authentication) and two SonicWall SMA1000 flaws. One addition, an OS command-injection bug in Kestra, was flagged after Microsoft traced it to a threat actor who used it in late June to establish a reverse shell, map Docker container environments, evade defenses and deploy a cryptocurrency miner. Federal civilian agencies face mandated remediation deadlines under CISA’s binding operational directive once a flaw lands in the catalog. Source: The Hacker News Verified: True

HPE shipped patches for 34 CVEs in its Aruba Networking ArubaOS-CX switch platform, including a cluster of critical-severity issues collectively tracked as CVE-2026-73749 with a CVSS score of 9.8. HPE says an unauthenticated attacker who sends malformed packets to an affected service could achieve remote code execution with elevated privileges, and it is urging customers to restrict CLI and web-management interfaces to dedicated network segments with firewall controls in place of relying on patching alone. The company says it has not observed active exploitation of the flaws so far, but the severity and network-facing exposure make rapid patching a priority for enterprises running AOS-CX in production. Source: SecurityWeek Verified: True

Enterprise Infrastructure

Nvidia confirmed a definitive agreement to acquire Hugging Face for $12.93 billion, roughly $11.9 billion payable to stockholders plus up to $1 billion in employee retention equity, in what would be Nvidia’s largest acquisition to date. Hugging Face’s platform hosts more than 3 million AI models, 1 million applications and 500,000 datasets used by over 18 million developers, and CEO Jensen Huang said it will stay an open platform where developers can pick any model, framework, cloud or compute provider without Nvidia hardware being required. The deal is expected to close in the first half of 2027 pending regulatory approval, and Nvidia has committed to preserving Hugging Face’s existing practice of supporting rival silicon vendors. Source: TechCrunch Verified: True

Equinix said it will launch Equinix Inference Exchange, a distributed AI inference service built with Nvidia’s validated Enterprise Reference Architectures and Together AI’s inference platform supporting more than 200 open-source models, delivered through Equinix’s data centers and Fabric interconnection. The service is aimed at enterprises trying to move from AI experimentation to production without building their own inference infrastructure, giving them low-latency connectivity to clouds, networks and AI providers across Equinix’s footprint of more than 280 data centers in 77 metros. The companies announced the partnership at Equinix’s inaugural Equinix Horizon event, with a target launch in the first quarter of 2027. Source: Equinix Verified: True

Policy & Regulation

The FTC and attorneys general from 22 states sued Amazon, alleging the company secretly converted its advertising auctions into de facto first-price auctions by adding an undisclosed “soft reserve price” surcharge in 2019 that it never told advertisers about. The complaint cites internal Amazon documents referring to “invented auction participants” and “proxy 2nd price” bids, and says the share of Sponsored Products advertisers paying their full winning bid climbed from roughly 30-40% in 2021 to about 80% by 2024, extracting tens of billions of dollars from more than a million brands and sellers. Amazon disputes the allegations and says its auction systems saved advertisers $8 billion between 2021 and 2025. Source: Federal Trade Commission Verified: True

California lawmakers closed out their 2026 legislative session near midnight on August 31 having given final approval to roughly 30 AI-related bills, covering chatbot safety, student privacy, AI-auditor registration, workplace surveillance limits, healthcare uses of AI and digital-replica rules. Governor Newsom now has until September 30, 2026 to sign or veto each measure individually, meaning the state’s next wave of AI regulation will be decided bill-by-bill over the coming weeks rather than through a single omnibus law. The volume of legislation underscores California’s continued role as a bellwether for state-level AI policy after SB 243’s earlier chatbot-safety rules took effect this year. Source: Transparency Coalition Verified: True