California Mandates AI Auditors as Google Pours €13B Into Finland

AI & Machine Learning

Sakana AI split its Fugu orchestration model into two variants, Fugu Max and Fugu Ultra v2, roughly eleven weeks after the original Fugu launch. Rather than a single foundation model, Fugu is a trained router that dynamically hands each step of a task to the cheapest capable model from a pool that includes Nvidia’s Nemotron family and other open-weight systems, then stitches the results together behind one OpenAI-compatible API. Fugu Max targets the best output per dollar at $2 per million input tokens and $6 per million output tokens, which Sakana says undercuts Sonnet 5 and Kimi K3 by 40 to 60 percent, while Fugu Ultra v2 chases maximum capability on hard, multi-step problems at $5 input and $30 output per million tokens. On the Chartography visual-reasoning benchmark, Fugu Ultra v2 scored 48.3 against Opus 5’s 27.3, achieved without any of the current flagship models from Anthropic or OpenAI in its routing pool. Both models are live immediately, with Sakana arguing that future gains come less from bigger single models and more from systems that know which machinery to deploy for a given task. Source: Sakana AI Verified: True

DeepSeek released V4.1-Flash, the smallest model in a new architecture family, as open weights on Hugging Face under an MIT license just as the Beijing-based lab said it would. The mixture-of-experts model carries a 552-billion-parameter backbone, nearly double the 284 billion in the prior V4-Flash, uses a causal encoder-decoder design, and supports up to a one-million-token context window with native multimodal input. DeepSeek said internal and third-party testing put the smaller model ahead of its own flagship V4-Pro on performance, cost, speed and total runtime, an unusual claim for a “Flash” tier model to make against its full-size sibling. Starting September 14, API requests addressed to V4-Pro will be silently rerouted to V4.1-Flash and billed at the cheaper model’s rate until a V4.1-Pro version ships, effectively retiring the current flagship without an announcement. The release extends a pattern of Chinese labs undercutting Western pricing while giving away weights that rivals charge a premium to license. Source: SiliconANGLE Verified: True

Consumer Hardware

LG used IFA 2026 in Berlin to unveil seven new home appliances built around its “Fit & Max” design philosophy and an expanded AI Home ecosystem. The headline device is a next-generation SIGNATURE refrigerator with LLM-based conversational AI for voice control, a 6.8-inch touchscreen and an “AI Fresh” mode that adjusts cooling zones based on what it detects inside. LG’s Fit & Max washer claims up to 70 percent less energy use than the European A-grade minimum thanks to a direct-drive motor, while a companion French-door refrigerator adds a “Zero Clearance” design meant to sit flush against surrounding cabinetry, and a 25-inch WashTower combines washer and dryer into a single body sized for European kitchens. The lineup also includes a downdraft cooking hood with touch-free motion controls and an AirDew hair dryer using LG’s “Triple Airfluidics” airflow technology. LG framed the showcase as a bet that AI’s biggest near-term consumer payoff is not chatbots but appliances that quietly manage energy and food waste in the background. Source: LG Newsroom Verified: True

Cybersecurity

SAP patched a maximum-severity, CVSS 10.0 vulnerability nicknamed OVERPASS in its kernel’s Extended Passport Processing component, used for tracing across S/4HANA, ERP, Business Suite and NetWeaver. Onapsis Research Labs, which discovered the flaw and reported it to SAP, found that missing boundary checks during deserialization of EPP data let an unauthenticated attacker trigger memory corruption when a new session opens across any of several communication protocols. Because authorization checks run only after the vulnerable code path is reached, an attacker can bypass authentication entirely to run arbitrary operating-system commands, pull database credentials and password hashes, hijack live user sessions, and modify SAP binaries. SAP shipped a single kernel patch, Security Note 3747649, covering kernel versions from 7.22 through 9.20, and Onapsis is urging emergency patching even though no active exploitation has been confirmed yet. Given SAP’s role running back-office systems for a large share of the world’s largest companies, a bug this severe and this easy to reach is expected to draw fast attention from ransomware groups. Source: SecurityWeek Verified: True

Check Point disclosed and same-day patched two 9.8-severity vulnerabilities in the VPN components of its Quantum Security Gateway, Security Management Server and Spark Firewall product lines. CVE-2026-85102 stems from improper certificate trust validation during VPN negotiation, while CVE-2026-85103 is a heap-based buffer overflow in how the software decodes ASN.1 structures inside VPN certificates; either can let an unauthenticated attacker execute code on the affected appliance. Check Point’s own research team found both flaws internally and says it has seen no evidence of active exploitation or public proof-of-concept code, but some customers reported delayed patch availability and broken download links as fixes rolled out. Affected releases include R82.10 with Jumbo Hotfix Take 43 or earlier, R82 with Take 125 or earlier, and R81.20 with Take 165 or earlier, with remediation available via automatic Live Patch rollout or manual hotfix installation. VPN gateways sit at the network edge, so unauthenticated RCE bugs here are a favorite target for opportunistic scanning within days of disclosure. Source: The Hacker News Verified: True

Enterprise Infrastructure

Rackspace Technology joined Nvidia’s Cloud Partner Program and introduced the “Institutional Sovereign Pod,” a governed AI infrastructure package combining Nvidia Blackwell accelerators with Palantir’s software stack for regulated enterprises and governments. The offering targets organizations that need to deploy AI at scale while keeping full control over their data, models and infrastructure, spanning compute, networking, visualization and enterprise software layers under a single accountable operator rather than a patchwork of vendors. Rackspace’s pitch is that as enterprise AI shifts from pilots to production infrastructure, someone needs to own uptime, governance and sovereignty end to end, not just sell raw GPU capacity. The deal continues a broader trend of legacy managed-infrastructure providers repositioning themselves as the operational layer for AI deployments that hyperscalers and chipmakers cannot service directly for sovereignty-sensitive customers. Source: Rackspace Verified: True

Google committed at least €13 billion to digital infrastructure in Finland over 2027 and 2028, its largest single investment in Europe, building three new data centers and expanding an existing one in Hamina while adding sites in Kajaani, Muhos and Vaala. Finnish Prime Minister Petteri Orpo called the investment “historic,” and Google separately signed a 22-year power purchase agreement with utility Fortum covering up to half the output of the Loviisa nuclear plant to help run the facilities on clean power. Google estimates the construction phase alone will support more than 37,000 jobs and add €3.6 billion annually to Finland’s GDP, with roughly 7,000 direct and indirect jobs continuing once the centers are operational and serving Gemini and other Google services. The company is also funding local clean-energy, biodiversity and workforce-development programs tied to the buildout. The move underscores how the AI infrastructure race is now as much about securing power contracts and land as it is about chips. Source: Google Verified: True

Policy & Regulation

California Governor Gavin Newsom signed two bills establishing what his office calls the first state framework for independent AI auditing. Senate Bill 813, from Senator Jerry McNerney, creates a certification process for independent verification organizations that can formally assess AI systems and models for compliance with state law, while Assembly Bill 1405, from Assemblymember Rebecca Bauer-Kahan, sets up a state registry of AI auditors with standards for their independence, transparency and integrity. The bills give California regulators and companies a state-sanctioned pool of third-party evaluators rather than relying solely on self-reported compliance, an approach Bauer-Kahan said is “essential to ensuring AI is safe for our communities.” Newsom paired the signings with a public call for federal AI legislation, arguing that a state-by-state patchwork is not a substitute for national rules. The move builds on California’s broader push this year to regulate frontier AI systems and follows a summer in which the state also enacted chatbot safety and child-protection measures. Source: Office of Governor Gavin Newsom Verified: True