Google Dodges Ad-Tech Breakup as US Pushes G20 to Ease AI Rules

AI & Machine Learning

Google released Gemini 3.8 Flash on September 2, calling it its most intelligent workhorse model and pairing it with a defense-only variant, Gemini 3.8 Flash Cyber. Google says 3.8 Flash outperforms larger frontier models on DeepSWE v1.1, a long-horizon software-engineering benchmark, and scores 54.9% on HLE-Verified for multi-step STEM and professional reasoning. Introductory pricing holds at $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026, and the model is available via the Gemini API, Android Studio, Gemini Enterprise and consumer AI Pro and Ultra plans. The Cyber variant claims a success rate above 70% on real-world vulnerability discovery across 20 languages and 47.2% pass@1 on the CWE-Bench automated-patching test, with Chrome Security reporting 2.6 times more correct patches than commercial alternatives. Google is restricting 3.8 Flash Cyber to vetted Fairwind participants — governments, critical-infrastructure operators and software maintainers — mirroring the gated-release approach rivals have adopted for offensive-capable models. The mid-tier Flash class is now where the coding-agent competition is fiercest, because cost per task matters more than peak capability for production deployments. Source: Google Verified: True

AI-safety researchers raised alarms over a reasoning method OpenAI is using in its forthcoming Astra model, known as recurrent depth or opaque recurrence, which lets a model iterate on a query through loops of internal computation instead of writing out step-by-step chain-of-thought text. The concern is monitorability: legible chain-of-thought has become a key tool for catching misaligned or deceptive behavior, and reasoning that happens in hidden latent space leaves far fewer traces to inspect. Staff at Redwood Research warned that scaling the technique could destroy researchers’ ability to monitor a model’s chain-of-thought, and that the natural next step is a system that reasons entirely in latent space. OpenAI says Astra’s use of the method is currently limited and that it still plans legible chain-of-thought monitoring, and both Anthropic and Google DeepMind are reportedly weighing similar approaches. The episode captures a widening tension between the capability gains that novel architectures unlock and the interpretability commitments the major labs made publicly in 2025. Source: TechCrunch Verified: True

Consumer Hardware

GoPro agreed to be acquired by Starman Optical, a US optical-transceiver and photonics manufacturer, in a roughly $285 million deal announced September 1 that values the action-camera maker at about $1.14 per share. GoPro shareholders will hold around 10% of the combined company, which will remain publicly traded, and the transaction clears roughly $92 million of GoPro debt. Starman wants to combine its photonics manufacturing with GoPro’s consumer optics and imaging expertise and expand into commercial and defense markets, framing advanced imaging as strategically important to AI and national security. The deal follows a brutal stretch for GoPro, which warned of possible insolvency in June and took a $20 million personal loan from founder Nick Woodman to keep operating. It is expected to close by the end of 2026. The outcome caps GoPro’s long slide from 2014 IPO darling to a distressed asset absorbed mainly for its lens and sensor know-how. Source: TechCrunch Verified: True

Mukesh Ambani’s Reliance Jio expanded JioPC, a cloud-PC service that streams a full virtual desktop from Jio’s data centers to any existing computer or TV in India over a broadband connection. The service turns machines up to eight years old into usable systems with up to eight virtual CPUs, 16GB of RAM and 1TB of storage, and it is now open to all Indian internet users rather than only Jio broadband subscribers. Pricing starts at about 1,000 rupees (around $11) for two months, with annual plans near 4,000 to 5,000 rupees (about $42 to $53). Analysts say it could create a low-cost computing tier for students and households priced out of new AI-capable PCs, as India’s PC replacement cycles stretch to five or six years. Reliance is betting that subscription access to remote compute, not hardware ownership, is the model for the next wave of Indian PC users. It is one of the most aggressive attempts yet to sell desktop computing as a utility. Source: TechCrunch Verified: True

Cybersecurity

SonicWall confirmed active exploitation of two vulnerabilities in its SMA1000 secure-access appliances and urged customers to install emergency hotfixes immediately. The more serious flaw, CVE-2026-83548, is a maximum-severity server-side request forgery in the Appliance WorkPlace interface that an unauthenticated attacker can chain toward remote command execution; CVE-2026-83549 lets an attacker who already has admin access run arbitrary operating-system commands from the management console. SonicWall’s product security team said it investigated a case indicating the flaws were being exploited in the wild, but did not detail the attacks or name a threat actor. Affected models include the SMA1000 6210, 7210 and 8200v; the SMA 100 series and standalone SSL-VPN products are not impacted. Shadowserver counted more than 400 internet-exposed SMA1000 appliances worldwide. Edge access gateways stay a favored target because a single unauthenticated bug exposes the internal network behind them. Source: BleepingComputer Verified: True

A dark-web site called Nexus began publishing what appears to be a massive trove of identity documents linked to IDScan, a Louisiana-based identity-verification provider, in what researchers are describing as one of the largest ID-document breaches on record. The collection reportedly holds more than 150 million US and Canadian driver’s licenses and passports, complete with the selfie photos submitted for verification, and the operators claimed to be adding roughly 500,000 new records a day through near-real-time access to a verification pipeline. Security journalist Brian Krebs confirmed his own license was in the data, and an ID photo said to belong to US Defense Secretary Pete Hegseth also appeared. IDScan said it was investigating, the FBI’s New Orleans field office opened a probe, and the Nexus site went offline shortly after the reports surfaced. The incident underscores how identity-verification vendors have become concentrated single points of failure, holding government-issued IDs for millions of people. Companies that outsource “are you real” checks inherit the breach exposure of whichever vendor they picked. Source: TechCrunch Verified: True

An international law-enforcement operation disrupted the infrastructure behind Sality, a peer-to-peer malware network that has persisted since 2003, seizing domains in the US and Europe with support from the Justice Department, FBI, Europol and Eurojust. CrowdStrike’s threat-intelligence team ran a sinkhole operation that isolated infected machines by targeting the super-peer nodes that relay commands across the botnet’s decentralized structure. Investigators said that in recent years the network mainly pushed clipboard-hijacking malware that swaps a victim’s copied cryptocurrency wallet address for one controlled by the operator. Authorities in Bulgaria, Hungary and Romania seized additional domains, and CrowdStrike attributes the operation to a Russia-based group it tracks as Salty Spider. Sality’s longevity shows how hard fully peer-to-peer botnets are to dismantle, since there is no central server to seize. The action degrades the network rather than eliminating it. Source: BleepingComputer Verified: True

Enterprise Infrastructure

Google signed a power-purchase agreement for 396 MW of enhanced geothermal electricity from Fervo Energy’s Cape Station project in Beaver County, Utah — its largest geothermal commitment to date and a foundational supply for a possible Google data center in the state. Deliveries are structured as four 99 MW tranches starting in the third quarter of 2028 under a 15-year term, with an option for Google to add roughly 600 MW by June 2030 for close to a full gigawatt. A broader framework agreement between the two companies covers development of up to 3 GW of geothermal capacity through 2033. Enhanced geothermal, which uses oil-and-gas drilling techniques to tap heat from hot dry rock, has moved from pilot to utility-scale procurement in about three years, and Fervo went public in May. The deal is another sign that hyperscalers see firm, carbon-free baseload power — not just intermittent renewables — as the binding constraint on AI data-center growth. Google had signed a contested natural-gas supply deal earlier in 2026. Source: TechCrunch Verified: True

Palo Alto Networks acquired Console, a two-year-old startup whose AI agents automate routine IT help-desk work such as password resets and application-access grants, for about $500 million in cash and stock. Console had raised only $29 million and was valued at roughly $157 million in its last round, making the price a steep step-up; its customers included Ramp, Flock Safety and Scale AI. Palo Alto plans to fold the technology into Cortex, its security-operations platform, so analysts can resolve alerts through natural-language instructions rather than manual playbooks. It is the company’s seventh acquisition of 2026 and fits a wider pattern of security vendors buying agentic-automation capability rather than building it in house. The deal leaves Sequoia-backed Serval as the most prominent independent startup in AI-driven IT service automation. Enterprise agent tooling is consolidating into the large platforms well before the category has matured. Source: TechCrunch Verified: True

Policy & Regulation

US District Judge Leonie Brinkema ruled that Google will not have to sell any part of its advertising-technology business, rejecting the Justice Department’s push to force a divestiture of its ad exchange and publisher ad server. Instead she ordered behavioral remedies that require Google to change how the ad-tech stack operates so competitors can gain access, with the detailed opinion staying sealed for 14 days to allow redactions. The decision follows Brinkema’s April 2025 finding that Google illegally monopolized parts of the online-advertising market. It echoes the outcome of the separate search-monopoly case, where a judge also declined to break up Google and imposed conduct restrictions instead. Google welcomed the rejection of a breakup, while the DOJ had argued that only structural separation could restore competition. Courts are now consistently choosing behavioral fixes over breakups in Big Tech antitrust cases, a pattern with major implications for pending and future actions. Source: TechCrunch Verified: True

The US used a G20 digital-ministers meeting in Chapel Hill, North Carolina, to press other governments toward a light-touch approach to AI rules, circulating a document it calls the Carolina Principles. The framework urges countries to reserve new regulation for genuinely novel problems, apply existing sector-specific laws rather than stand up dedicated AI regulators, direct public money toward foundational research, and coordinate closely with industry. White House science and technology chief Michael Kratsios co-hosted the meeting; reporting says China has signed on, and the principles are expected to go to G20 leaders for formal endorsement at a December summit in Florida. The pitch aligns closely with the interests of large US AI companies, which want to avoid a patchwork of stricter foreign rules. The EU is moving the other way, advancing new AI legislation, setting up a clear transatlantic divergence. It is the sharpest articulation yet of the administration’s effort to export a deregulatory AI posture. Source: Al Jazeera Verified: True