AI News
Daily Tech Reader - AI Edition
Podcast 🎧 • Video 📽 • Short 📽
Apple
Apple
Apple Patches Critical ImageIO Zero-Click Flaw Across iOS, iPadOS, and macOS
Apple's August 17 security updates fix CVE-2026-65346, an integer-overflow bug in ImageIO discovered by Meta's Red Team that could enable arbitrary code execution via malicious images. Experts warn AI-accelerated exploit development is compressing the window between disclosure and weaponization.
Sources: The Register · Tech Startups
CISA Gives Federal Agencies Three Days to Patch Ray AI Framework RCE Bug
CISA added a critical CVSS 9.4 remote-code-execution flaw in Ray — the open-source AI framework used by Apple, Amazon, and OpenAI — to its Known Exploited Vulnerabilities catalog, setting an August 20 deadline. The ShadowRay 2.0 campaign is already converting compromised GPU clusters into cryptomining botnets.
Sources: The Hacker News · AI Weekly
Microsoft
Microsoft Finally Patches One-Click Copilot Data Theft Flaw — Eight Months After Discovery
Microsoft shipped a fix for CVE-2026-24301 (dubbed CoSnitch), a critical Copilot vulnerability that chained a hidden URL parameter, persistent memory poisoning, and Copilot's built-in URL fetch to silently exfiltrate Gmail, Drive, and Calendar data. Varonis reported the flaw in December 2024; the full patch took roughly eight months.
Sources: CSO Online · AI Weekly · Computerworld
Wiz Red Team: GitHub Copilot Autofix Patch Opened Shell-Injection Hole in Snowflake Repo
A Copilot Autofix-generated patch to Snowflake's connector repo replaced safe input handling with raw string interpolation, creating a shell-injection vulnerability exploited within five days. The flaw let an unauthenticated attacker exfiltrate a Jira token granting read access to engineering, security, and bug bounty projects.
Sources: Wiz · AI Weekly
Guardian: Microsoft's 2.2 Million AI Chips Fall Far Short of Stated Data Center Capacity
An investigation citing internal documents found Microsoft has 2.2 million AI chips installed globally — well below the 6.4 million GPUs experts say its $280B data center buildout should support. Microsoft disputed the math; MSFT shares fell roughly 3%.
Sources: The Guardian · AI Weekly
Claude
Anthropic Signs $250M Inference Chip Deal With UK Startup Fractile — Valuation Jumps to $6.5B
Oxford spinout Fractile is raising ~$600M at a $6.5B valuation — a sixfold jump from May — after an initial agreement to supply Anthropic with roughly $250M of its SRAM-based inference chips, with both sides signaling intent to expand. Chips are not expected to be production-ready until 2027.
Sources: The Next Web · AI Weekly
Claude Designed Protein Binders for 14 of 15 Targets at More Than Double the Industry Success Rate
Anthropic published lab-validated results showing Claude (Mythos Preview and Opus 4.8) achieved a 22–35% binder success rate across 15 protein targets, compared with a 10–15% industry baseline. Opus 5 also processed raw NMR and LC-MS data in under 25 minutes with purity within 0.1% of lab readings.
Sources: Anthropic · AI Weekly
Anthropic Story Volume Surges 45% This Week as Inference, Chips, and Safety Dominate Coverage
AI Weekly's entity tracker shows Anthropic at 59 tracked stories this week, up 45% from last week's 22 — the fastest riser among all 113 tracked AI entities. The surge is driven by the Fractile chip deal, protein design results, and ongoing EU AI Act watermarking coverage.
Sources: AI Weekly
ChatGPT
OpenAI Bakes 20% Compute Overhead Into Astra Inference for Mandatory Safety Monitoring
OpenAI now estimates safety monitoring overhead at roughly 20% of inference compute for GPT-5.6 Sol-class and higher models and all Astra inference, following its "Pacing model development in an era of cyber-critical capabilities" policy update. The company says the costs are internal research spend and will not be passed to customers — for now.
Sources: The Register · AI Weekly
OpenAI CFO Commits to 2027 IPO — or Sooner If Business Continues to "Inflect"
At an all-hands meeting, CFO Sarah Friar told employees OpenAI "will be a public company in 2027," with an earlier listing possible if growth accelerates. The commitment aligns with prior reporting that Friar was pushing for a 2027 date against Sam Altman's preference for a 2026 debut.
Sources: CNBC · AI Weekly
OpenAI Pilots Private Safety Processing — Zero-Retention Monitoring for Paid API Customers
OpenAI is testing Private Safety Processing with early API customers, a system designed to flag misuse and misaligned agent behavior while preserving zero-data-retention guarantees. Anthropic's competing safety practices currently require data logging in exchange for higher rate limits.
Sources: Bloomberg · Axios · AI Weekly
Gemini
Google Gives U.S. College Students a Free Year of Gemini Pro — International Students Get Gemini Plus
Eligible U.S. students get 12 months of Google AI Pro free (normally $19.99/month); international students in 140+ markets receive a free year of the AI Plus plan (normally $4.99/month). Google is also launching a Student Hub with diagnostic quizzes, interactive 3D visualizations, and Deep Research inside Gemini Live through December 31.
Sources: Google · AI Weekly
Google Grants Marvell a $12.2B Warrant in Custom AI Chip Deal Covering Inference, Storage, and Networking
Google gained the right to purchase up to ~59 million Marvell shares at $206.58 apiece in exchange for a custom-chip supply agreement covering AI inference accelerators, memory-interface controllers, and near-memory computing. Marvell stock jumped more than 10% on the news; rival Broadcom fell roughly 3%.
Sources: Bloomberg · AI Weekly
China AI
Unitree Robotics Surges 629% on Shanghai Debut — Valuation Briefly Hits $66 Billion
Humanoid robot maker Unitree opened at 1,100 yuan, up 629% from its IPO price, briefly reaching a $66B valuation before settling around 883 yuan in the largest first-day gain for a Star Market listing this year. The retail tranche was oversubscribed more than 5,500 times; DeepSeek invested roughly 140 million yuan as a backer.
Sources: SCMP · AI Weekly · CNBC
Z.ai Releases GLM-5.3 Open-Weight Model Built for Coding, Cybersecurity, and Long-Horizon Agentic Tasks
Chinese AI lab Z.ai released GLM-5.3 via API at the same price as GLM-5.2, targeting advanced coding and defensive cybersecurity work at near-frontier capability. The open-weight release adds another downloadable model to China's growing arsenal as U.S. labs debate the security implications.
Sources: Crypto Integrated · WIRED
Big Story
Cerebras CS-4 Claims 30x GPU Inference Speed as AI Compute Race Expands Beyond Nvidia
Cerebras unveiled the CS-4, a rack-scale system built on three next-generation wafer-scale engines claiming up to 30x faster inference than GPU deployments and more than 1,000 tokens per second on 10-trillion-parameter models. First shipments begin this quarter as the inference market becomes AI's most contested battleground.
Sources: Cerebras · AI Weekly · Investor's Business Daily
Samsung Raises Advanced Foundry Prices 15% as AI Demand Fills Every Nanometer of Capacity
Samsung hiked prices for its 4nm and 5nm foundry processes by 10–15% and its 8nm line by nearly 10%, with its Pyeongtaek SF4 fab running at full capacity since late 2022. AI applications are now expected to top 30% of Samsung foundry revenue as TSMC overflow demand fills the gap.
Sources: Reuters · AI Weekly
Amazon Makes Alexa+ Free on All Fire TV Devices — No Prime Subscription Required
Amazon is auto-upgrading all US Fire TV Sticks, Cubes, and Ember TVs with Alexa+, eliminating the previous $19.99/month non-Prime fee with no new app or subscription needed. The company says Alexa+ users have nearly twice as many assistant conversations as users of the legacy Alexa.
Sources: TechCrunch · AI Weekly