Altman’s AI Slowdown U-Turn Meets California’s New Chatbot Safety Law

AI & Machine Learning

OpenAI CEO Sam Altman told employees this week that the company is open to pacing the development of its most advanced AI systems and hopes rival labs will follow, a sharp reversal after years of arguing that slowing down AI was both impossible and unwise. The shift follows mounting internal alarm over autonomous agents behaving unpredictably during recent evaluations, and comes days after a researcher who had worked at both OpenAI and Anthropic publicly accused the two companies of “gambling with our lives” by racing toward more capable systems without matching safety controls. Altman reportedly discussed a slower cadence specifically for models trained beyond 10^26 floating-point operations, and pointed to voluntary commitments OpenAI has made with Anthropic and Google DeepMind to share safety incidents with the U.S. AI Safety Institute. Whether competing labs actually pull back remains an open question, since a coordinated slowdown only works if rivals do not use the pause to pull ahead. Source: Bloomberg Verified: True

Sakana AI released two new orchestration models, Fugu Max and Fugu Ultra v2, replacing its original Fugu router with a split lineup aimed at cutting multi-agent costs while pushing peak capability higher. Rather than a single foundation model, Fugu is a learned router that dispatches work across a pool of underlying models through one API, and Ultra v2 targets the hardest end of that spectrum, posting 48.3 on the Chartography benchmark versus 27.3 for Anthropic’s Opus 5 and 74.3 on DeepSWE. Sakana priced Ultra v2 at $5 per million input tokens and $30 per million output tokens, roughly 60% cheaper than running the prior Fugu Ultra for comparable workloads, with a doubled rate for contexts above 272K tokens. Both models are live now through Sakana’s OpenAI-compatible API, and the company says switching from an earlier Fugu version requires only a one-line model ID change rather than a full SDK migration. Source: MarkTechPost Verified: True

Consumer Hardware

Xiaomi unveiled the 18 Fold, a “mid-size” foldable built to undercut Samsung’s Galaxy Fold 8 and get ahead of Apple’s own foldable iPhone before it ships. The device pairs a 7.58-inch inner display with a 5.38-inch cover screen, both using a taller A4 aspect ratio, and at 219 grams it is the lightest foldable Xiaomi has made, powered by the company’s in-house XRING O3 chip. Co-developed with Leica, the 18 Fold carries a 200MP main camera, a silicon-carbon 6,000mAh battery with 67W wired charging, up to 16GB of RAM, and ships with Android 17, starting at roughly $1,640 in China. Hands-on impressions from IFA 2026 in Berlin praised the phone’s flat, ultra-thin-bezel design and asymmetric front layout as a distinct alternative to Samsung’s more conventional foldable shape. Source: Engadget Verified: True

Cybersecurity

GitLab is warning customers to immediately patch CVE-2026-85706, a maximum-severity path traversal flaw in its repository commits API that lets an unauthenticated attacker read arbitrary files, including credentials, SSH keys and configuration data, from an affected GitLab server. GitLab attributed the bug to “improper path confinement and missing authentication enforcement” and shipped fixes in versions 19.1.8, 19.2.6 and 19.3.2 for Community and Enterprise Edition installs running 18.7 through 19.3.1. Researchers at watchTowr observed active, internet-wide exploitation attempts beginning at 06:00 UTC on September 11, just hours after the patch and advisory went public, and CISA added the flaw to its Known Exploited Vulnerabilities catalog the same day. Federal civilian agencies now have until September 14 to patch, and any self-managed GitLab instance exposed to the internet should be treated as at immediate risk. Source: The Hacker News Verified: True

Separately, CISA added three actively exploited flaws in Cisco, Citrix and Fortinet products to its Known Exploited Vulnerabilities catalog, setting a September 12 patch deadline for federal agencies. The Cisco bug, CVE-2026-20079, is a maximum-severity authentication bypass in Secure Firewall Management Center’s web interface that lets an unauthenticated attacker execute script files and obtain root access, while CVE-2026-19490 lets attackers bypass authentication on Citrix NetScaler ADC and Gateway when configured as an AAA or Gateway virtual server. The third flaw, in Fortinet software, has been weaponized to deliver a Node.js remote access trojan CISA is tracking as PivotC2; researchers estimate more than 3,000 IP addresses were targeted and 178 devices infected in a campaign attributed to a financially motivated, Russian-speaking threat actor. The cluster of three simultaneous KEV additions underscores how quickly edge network and firewall appliances are becoming the preferred entry point for ransomware-adjacent groups. Source: The Hacker News Verified: True

Enterprise Infrastructure

Analog Devices agreed to acquire Alif Semiconductor for $1.35 billion in cash, plus an earn-out of up to $200 million, adding an AI-native edge processing platform aimed at what ADI calls “physical intelligence” for robotics, industrial and other real-world systems. Alif’s chips are designed to run AI inference directly on edge devices with tight power budgets, complementing ADI’s existing sensor and signal-processing business rather than competing with cloud-scale AI silicon from Nvidia or AMD. The deal has been approved by both companies’ boards and still needs to clear Hart-Scott-Rodino antitrust review, which ADI expects to complete by the end of 2026. It is one of the larger semiconductor acquisitions of the year and signals growing appetite among established chipmakers to buy their way into edge-AI processing rather than build it from scratch. Source: Analog Devices Verified: True

Italian app developer Bending Spoons agreed to acquire collaborative whiteboard company Miro at an enterprise value of $1.355 billion, extending its recent run of buying established software brands including Evernote and Vimeo. Miro serves more than 250,000 organizations and has grown to roughly $600 million in annual recurring revenue, much of it from enterprise teams using the platform for visual collaboration and, increasingly, AI-assisted workflow tools layered on top of its boards. The all-cash transaction is expected to close in the fourth quarter of 2026 pending regulatory approval, with Bending Spoons planning to keep Miro operating as a standalone product. The acquisition adds a sizable B2B SaaS asset to a buyer better known for consumer apps, continuing a broader wave of consolidation among mid-size collaboration and productivity software companies. Source: Bending Spoons Verified: True

Policy & Regulation

California Governor Gavin Newsom signed 13 bills on September 10 establishing what his office called the nation’s strongest child-safety protections covering AI chatbots and social media, headlined by SB 1119, known as Adam’s Law. The law requires companies operating AI companion chatbots to run crisis protocols that trigger when a minor’s conversation suggests suicidal ideation or self-harm, verify users’ ages, provide access to crisis resources, notify parents in qualifying situations, and complete independent child-safety audits and annual risk assessments. A companion measure, AB 1709, bars social platforms from offering addictive design features like autoplay and algorithmic feeds to users under 16, while SB 1276 expands child sexual exploitation statutes to cover AI-generated or digitally altered depictions of minors, with penalties up to $250,000 per violation. Adam’s Law is named for Adam Raine, a California teenager whose 2025 death by suicide led to a wrongful-death lawsuit against OpenAI over his extensive conversations with ChatGPT about self-harm, and it makes California the first state with a comprehensive regulatory framework specifically for AI companion chatbots and children. Source: Office of Governor Gavin Newsom Verified: True