Google unveiled scalable liquid cooling for AI data centers at Hot Chips 2025, emphasizing water’s high thermal conductivity and maintenance protocols to meet AI heat demands, alongside new security silicon from Microsoft and critical patches for Citrix, Docker, and Google’s Gemini AI models. Meanwhile, Nevada recovers from a cyberattack, and legislation efforts target ghost job postings and AI safety.
▶️ Internet Infrastructure
Google Unveils Scalable Liquid Cooling for AI Data Center TPUs at Hot Chips 2025
Google’s Hot Chips 2025 talk details scalable liquid cooling for datacenter TPUs, utilizing water’s high thermal conductivity, split-flow cold plates, bare-die chips, and maintenance protocols to handle AI power and heat demands.
- Google presented datacenter-scale liquid cooling for TPUs at Hot Chips 2025, emphasizing high thermal conductivity of water (~4000x air) for AI workloads.
- The cooling system uses racks with six Coolant Distribution Units (CDUs), flexible hoses, quick disconnects, and a series loop of TPU chips with split-flow cold plates; TPUv4 chips are bare-die, 1.6x higher power than TPUv3.
- Pump power in liquid cooling is less than 5% of fan power in air cooling; Google employs leak testing, alerting, scheduled maintenance, and zero-downtime protocols, contrasting with enthusiast PC water cooling challenges.
Over 19 Million Malware Apps Downloaded from Google Play with Advanced Evasion Techniques
Zscaler reports over 19 million downloads of malware-laden apps in Google Play, including updated Anatsa banking trojan variants with advanced evasion, highlighting security flaws in Google’s app vetting process.
- Over 19 million malware-infected apps were downloaded from Google Play Store, evading Google’s security scans, according to Zscaler ThreatLabz.
- 77 malicious apps were identified, many mimicking utilities or personalization tools, with some containing updated Anatsa banking trojan featuring keylogger, SMS interception, and anti-detection.
- The latest Anatsa version uses code chunk downloads with DES keys, alters app names, and employs obfuscation techniques like corrupted ZIP archives and dynamic JSON payloads to evade static analysis.
Citrix Patches Critical NetScaler Vulnerabilities Amid Active Exploits
Citrix patched three NetScaler bugs (CVE-2025-7775, CVE-2025-7776, CVE-2025-8424), with CVE-2025-7775 already exploited in the wild, enabling remote code execution and persistent access.
- Citrix released patches for three NetScaler vulnerabilities: CVE-2025-7775, CVE-2025-7776, and CVE-2025-8424
- CVE-2025-7775 is a pre-auth remote code execution memory overflow with a CVSS score of 9.2, exploited in the wild before patching
- Attackers used the zero-day flaws to deploy webshells and backdoors, with incident response recommended for affected organizations
Docker patches critical CVE-2025-9074 flaw exposing host systems via Docker API
Docker Desktop fixed a critical vulnerability (CVE-2025-9074, CVSS 9.3) that exposed the Docker Engine API on TCP port 2375, enabling containers to break isolation and control host systems, especially on Windows via WSL2.
- Docker patched CVE-2025-9074, a critical vulnerability scoring 9.3 CVSS, exposing Docker Engine API on “192.168.65.7:2375” without authentication.
- The flaw allowed containers to access the host system, mount drives, and execute arbitrary code, with Windows users gaining full read-write access via WSL2.
- The vulnerability was discovered by Felix Boulet and demonstrated by Philippe Dugre, who showed how trivial abuse was on macOS with three lines of Python code; Linux Docker Engine was unaffected.
Microsoft Unveils Custom Security Silicon with HSMs at Hot Chips 2025
Microsoft unveiled custom security silicon with integrated HSMs and Caliptra 2.0 RoT modules at Hot Chips 2025, enhancing Azure’s hardware security with disaggregated, tamper-resistant cryptographic components.
- Microsoft presented its new security silicon, including HSM and Caliptra 2.0 Root of Trust modules, at Hot Chips 2025.
- The integrated HSMs are optimized for AES and Private Key Encryption, hardened against physical and side-channel attacks, and use sealed anti-tamper packaging.
- Disaggregated HSM functionality is achieved via custom silicon, replacing traditional appliances, to improve scalability and reduce latency in Azure’s compute infrastructure.
Nevada Continues Recovery After Two-Day Cyberattack Disrupts State Services
Nevada’s government faces ongoing recovery two days after a cyberattack disabled websites, phone lines, and state offices, with no confirmed data breach and active incident response.
- Nevada experienced a two-day cyberattack starting August 24, disrupting state websites and services
- Official websites and Department of Motor Vehicles (DMV) offices remain closed; phone lines are intermittently down
- State officials rely on social media for updates; no evidence of personal data breach reported; recovery protocols are active
Nikkei and Asahi Sue Perplexity for Copyright Infringement and Data Violations
Nikkei and Asahi sued Perplexity for scraping articles, ignoring robots.txt, and disseminating content without authorization, seeking ¥4.4 billion in damages, amid broader copyright and AI data use disputes.
- Nikkei and Asahi filed a joint lawsuit on August 26, 2025, in Tokyo District Court against Perplexity for copyright infringement
- Allegations include unauthorized scraping of articles, ignoring robots.txt files, and serving content in response to user queries
- The suit seeks an injunction and ¥2.2 billion ($15 million) in damages per media firm; claims Perplexity violated exclusive rights and cited incorrect information, breaching Japanese copyright law
VMware Customers Shift to Nutanix Amid Licensing Changes and Cloud Modernization
Many VMware customers are choosing Nutanix as an alternative amid Broadcom’s post-acquisition licensing issues, benefiting from flexible deployment options, AI readiness, transparent pricing, and proven ROI.
- VMware customers are shifting to Nutanix following Broadcom’s acquisition of VMware and subsequent licensing and pricing changes.
- Nutanix offers a unified cloud platform supporting virtual machines and containers across edge, datacenter, and public clouds, enabling modernization for AI and cloud-native development.
- Customers migrating from VMware fall into two categories: modernizers adopting containerized and AI capabilities, and gradualists supporting hybrid multicloud infrastructure with partnerships with Dell, Pure Storage, NVIDIA, and Hugging Face; large-scale migrations include Computershare’s 24,000 VMs.
SSA Whistleblower Alleges Unsecured DOGE Cloud Copy of Critical Database
A SSA whistleblower alleges DOGE duplicated the NUMIDENT database in an unsecured cloud, risking identity theft and benefits loss, with potential government re-issuance costs.
- Whistleblower complaint filed by SSA’s chief data officer alleges DOGE duplicated the NUMIDENT database in an unmanaged cloud environment lacking security controls
- The database copy was made in June, with only two DOGE employees having administrator access, bypassing SSA protocols
- The complaint warns that malicious access could lead to identity theft, loss of benefits, and costly re-issuance of Social Security Numbers
▶️ Open Source
Framework Laptop 16 2025: Modular Power with Upgradeable GPUs and Open Source Design
Framework Laptop 16, released in 2025, offers modular, upgradeable hardware including NVIDIA® GeForce RTX™ 5070 and AMD Radeon™ RX 7700S GPUs, AMD Ryzen™ AI 300 Series CPUs, up to 10TB storage, and open source design files.
- Framework Laptop 16 features upgradeable graphics with NVIDIA® GeForce RTX™ 5070 and AMD Radeon™ RX 7700S modules, released in 2025.
- Equipped with AMD Ryzen™ AI 300 Series processors up to 12 cores and boost speeds of 5.1 GHz, supporting up to 96GB DDR5 RAM and 10TB storage.
- Supports open source ecosystem with design files on GitHub, modular components, and extensive customization options, including multiple input modules and expansion cards.
Custom Logitech MX Ergo Gets USB-C Charging and Silent Switch Upgrade
The author customized the Logitech MX Ergo by adding USB-C charging through PCB modifications, replacing switches with silent Huano and Omron models, and switching to SteerMouse software for better control.
- The author modified the Logitech MX Ergo (2017) to add USB-C charging, replace switches with silent Huano and Omron models, and upgrade software to SteerMouse.
- The USB-C port replacement involved reverse-engineering and custom PCB fabrication via PCBWay, costing approximately $45 for parts and materials.
- Switch replacements included desoldering and soldering Huano Silent switches ($2.99) for clicks and Omron B3F-1002 and Alps SKQGABE010 for scroll-wheel and other buttons.
OpenAI to Enhance ChatGPT’s Mental Health Detection and Safeguards
OpenAI will update ChatGPT to improve detection of mental distress cues and enhance safeguards after a lawsuit linked its AI to a teen’s suicide, aiming to prevent harmful interactions.
- OpenAI announced plans to update ChatGPT to better detect and respond to expressions of mental distress following a lawsuit alleging its role in a teenager’s suicide.
- The updates will include recognizing signs like sleep deprivation and advising users to rest, and strengthening safeguards around conversations about suicide.
- The company’s blog post emphasizes improving AI responses during prolonged conversations to prevent breakdowns in mental health support.
▶️ Management and Leadership
US invests $8.9B in Intel to secure chip supply amid geopolitical tensions
The U.S. government’s 10% stake in Intel aims to secure semiconductor supply chains amid geopolitical risks, addressing long-term industry challenges and Taiwan dependency, despite criticism over market distortions.
- U.S. government announced an 10% equity stake in Intel, valued at $8.9 billion, as part of industrial policy measures.
- Critics, including Scott Lincicome, argue this risks politicizing Intel, undermining fiduciary duties, and distorting market competition.
- The geopolitical context emphasizes Taiwan’s strategic importance for chip manufacturing, especially for AI and military applications, complicating U.S. dependency on TSMC and Samsung.
Verily Discontinues Medical Devices to Focus on AI and Data Strategy
Verily, Alphabet’s life sciences unit, laid off staff and shut down its devices program to focus on AI, data, and infrastructure, as part of a strategic shift since 2023.
- Verily laid off staff and discontinued its medical devices program, including projects like Study Watch, retinal camera, and Dexcom glucose sensor partnership.
- CEO Stephen Gillett cited strategic refocusing on AI, data, and infrastructure, with a shift away from manufacturing medical devices.
- The company has been streamlining since 2023, selling its insurance business and re-benchmarking employee salaries to healthcare industry standards.
Advocates Push for Legislation to Ban Ghost Job Postings and Ensure Transparency
Eric Thompson advocates for the Truth in Job Advertising and Accountability Act to ban ghost job postings, which 17% of Q2 2025 listings on Greenhouse represent, by requiring transparency, limiting duration, and imposing fines for violations.
- Eric Thompson, after losing his job in October 2024, led efforts to propose the Truth in Job Advertising and Accountability Act (TJAAA) to ban ghost job postings.
- Ghost jobs are listings posted without the intent to hire, comprising 17% of all job posts on Greenhouse in Q2 2025; companies use them to fill multiple vacancies, collect resumes, or signal growth.
- The TJAAA draft mandates transparent job posting details, limits posting duration to 90 days, requires a minimum four-day application window, and imposes a minimum $2,500 fine per violation for companies with over 50 employees; legislation aims to address the practice at the federal level due to multi-state and third-party platform complexities.
OAuth: Securing Third-Party Access with Authorization Codes and Tokens
OAuth, created in 2007, uses authorization codes and access tokens to securely grant third-party apps limited user data access, involving user consent, redirect URIs, and secure backend exchanges.
- OAuth was introduced in 2007 by Twitter to enable third-party apps to access user data securely without sharing passwords.
- The core mechanism is the access token, which grants limited, user-specific API access, obtained via an authorization code exchange.
- OAuth flow involves redirecting users to authorization servers, obtaining authorization codes, and exchanging them securely over HTTPS, avoiding insecure URL token transmission.
Apple Announces September 2025 Event with iPhone 17 and New Devices
Apple announced its September 9, 2025, event at 10:00 a.m. PDT at Apple Park, unveiling iPhone 17 series, new Apple Watch models, and upcoming software updates, streamed live.
- Apple announced its September 2025 event for September 9 at 10:00 a.m. PDT at Apple Park, Cupertino
- The event will unveil the iPhone 17 lineup, including an all-new ultra-thin iPhone 17 Air, iPhone 17, Pro, and Pro Max models
- Expected features include a new durable aluminum build, redesigned rear camera module, A19 or A19 Pro chips, and potentially ProMotion across all models; iPhone 17 Air will be the thinnest and lightest, with a 6.6-inch display, C1 modem, and single-lens camera
- Apple Watch Series 11, Watch Ultra 3, and Watch SE 3 will be announced simultaneously, alongside possible AirPods Pro 3, a new HomePod mini, and refreshed Apple TV 4K
- Apple will also announce launch dates for iOS 26, iPadOS 26, macOS Tahoe, and other software updates
- The event will be streamed live on Apple’s website, YouTube, and the Apple TV app, with coverage on MacRumors
Google to Mandate Developer Verification for All Android Apps Starting 2026
Google extends developer verification to all Android apps from September 2026, requiring registration, ID, and a $25 fee, aiming to enhance security amid malware concerns.
- Google will require all Android apps, including sideloaded ones, to undergo developer verification starting September 2026 in Brazil, Indonesia, Singapore, and Thailand, with a global rollout in 2027.
- Developers must create an Android Developer Console account with a one-time $25 fee, linked to a Google payment profile, and submit government-issued ID, verified phone number, and, for organizations, business registration and website verification.
- Each app package must be registered with a public key certificate, linked to the verified developer, without needing to upload the APK; existing Play Store developers are already verified.
Anthropic Launches Claude for Chrome Amid Security and Privacy Concerns
Anthropic’s Claude for Chrome preview allows AI-driven browsing for 1,000 paying users, highlighting security vulnerabilities, prompt injection risks, and ongoing mitigation efforts before general release.
- Anthropic offers a research preview of Claude for Chrome, limited to 1,000 paid subscribers at $100-$200/month, with safety warnings.
- The extension enables Claude to automate web browsing but introduces significant security and privacy risks due to broad permissions.
- Anthropic reports a prompt injection attack success rate of 23.6% without mitigations, reduced to 11.2% with safety measures; specific browser attacks reduced from 35.7% to 0%.
GSA Fast-Tracks AI Chatbots for Federal Use Through FedRAMP 20x Approval
The GSA is fast-tracking AI chatbots for federal use via FedRAMP 20x, enabling companies with MAS contracts to bypass approval hurdles, though meeting strict criteria remains challenging.
- GSA will prioritize authorization of cloud services with “conversational AI engines” for routine use, allowing them to bypass standard FedRAMP approval.
- AI products will use FedRAMP 20x, a program designed to streamline cloud software approval via automation, continuous monitoring, and standardization.
- Criteria for AI prioritization include enterprise security, data protection guarantees, demand from at least five CFO Act agencies or CIO Council recommendation, availability via Multiple Award Schedule (MAS), and compliance with FedRAMP 20x standards within two months.
- Companies like OpenAI, Google, and Anthropic have signed MAS contracts with GSA, potentially facilitating faster approval, though full commitment to FedRAMP standards remains unverified.
Microsoft to Enable OOBE Windows Updates via Intune Starting September 2025
Microsoft will enable administrators to install Windows quality updates during OOBE via Intune starting September 2025, improving initial setup update status for Windows 11 22H2+ devices.
- Starting September 2025, Windows administrators can deploy quality updates during Out of Box Experience (OOBE) via Intune or Windows Autopilot Enrollment Status Page (ESP).
- The feature applies to devices running Windows 11 22H2 or later, enrolled with Windows Autopilot, and imaged with the June 2025 update or later.
- Updates during OOBE are controlled by a policy setting, respect pause and deferral options, and are enabled by default for new ESP profiles; cannot be disabled if not using device ESP.
Trump Threatens Tariffs and Bans Over Digital Taxes on US Big Tech
Donald Trump threatened extra tariffs and tech export bans against nations regulating US Big Tech, citing discriminatory digital taxes and legislation, with potential restrictions on technology and chips.
- Donald Trump threatened to impose additional tariffs and export bans on countries regulating US Big Tech companies, citing digital taxes and regulations as discriminatory.
- He announced potential “substantial additional Tariffs” and export restrictions on advanced technology and chips, claiming to defend American tech firms.
- Trump criticized digital services taxes, digital legislation, and regulations, alleging they harm US companies and favor Chinese firms; he also highlighted the profitability and tax minimization schemes of US tech giants.
▶️ Technology
Google Launches Gemini 2.5 Flash Image for Advanced AI Image Generation
Google introduced Gemini 2.5 Flash Image, a high-quality image generation and editing model supporting multi-image fusion, character consistency, prompt-based transformations, and world knowledge, available via API and AI Studio for $0.039 per image.
- Gemini 2.5 Flash Image (aka nano-banana) launched on August 26, 2025, as a state-of-the-art image generation and editing model
- Enables blending multiple images, maintaining character consistency, targeted transformations via natural language, and leveraging Gemini’s world knowledge
- Available through Gemini API, Google AI Studio, and Vertex AI; priced at $30 per 1 million output tokens, with each image costing $0.039 (1290 tokens)
Google Gemini 2.5 Flash Image Launches with Advanced AI Image Editing Capabilities
Google’s Gemini 2.5 Flash Image introduces an AI image editor with fine control, multi-turn conversation capabilities, and state-of-the-art benchmark performance, aiming to compete with OpenAI’s image tools.
- Google announced Gemini 2.5 Flash Image, an AI image model enabling precise, natural language-based edits while maintaining facial and object consistency.
- The update, rolling out August 26, 2025, to Gemini app, API, Google AI Studio, and Vertex AI, improves visual quality and instruction-following.
- The model is state-of-the-art on benchmarks like LMArena, supports multi-turn editing, and is designed for consumer use cases such as home visualization.
Will Smith Concert Video Sparks AI-Generated Crowd Controversy
Real concert footage was enhanced with AI image-to-video models and YouTube’s unblurring technology, creating misleading visuals that appear AI-generated but are based on authentic crowd images.
- A short clip of Will Smith’s concert went viral amid accusations of AI-generated fake crowds and signs.
- The crowds in the video are confirmed to be real, sourced from multiple European tour performances, including the Positiv Festival in Orange, France.
- The video was manipulated through two levels: AI-generated short clips from real photos and YouTube’s AI-based post-processing, which sharpened and denoised the footage without using generative AI for the entire crowd.
Anthropic Tests Claude Chrome Extension to Boost AI Safety and Reduce Attacks
Anthropic is testing a Claude Chrome extension with 1,000 trusted users, reducing prompt injection attack success from 23.6% to 11.2% through safety measures, aiming to improve browser AI safety.
- Anthropic is piloting a Chrome extension for Claude, enabling direct browser actions for 1,000 Max plan users in controlled testing.
- Extensive adversarial prompt injection testing showed a 23.6% attack success rate without mitigations, reduced to 11.2% with current safety measures.
- Implemented defenses include site-level permissions, action confirmations, improved system prompts, blocking high-risk websites, and advanced classifiers; specific browser attack success rate reduced from 35.7% to 0%.
MVNOs Offer Affordable and Flexible Single-Line Plans on Major Networks
MVNOs like Mint Mobile, Visible, US Mobile, and Tello provide cost-effective single-line plans with comparable data and coverage, leveraging major networks and offering flexible, affordable options.
- MVNOs like Mint Mobile, Visible Wireless, US Mobile, and Tello offer more cost-effective single-line plans than major carriers (Verizon, AT&T, T-Mobile), which are better suited for multi-line accounts.
- Mint Mobile’s 12-month unlimited plan costs $15/month, provides unlimited talk, text, high-speed data (79-357 Mbps), and operates on T-Mobile’s network with deprioritization after 35GB.
- Visible Plus Plan offers truly unlimited high-speed data on Verizon’s network for $29/month, including unlimited hotspot data and coverage across LTE and 5G Ultra Wideband.
- US Mobile’s Unlimited Starter Plan allows network choice among Verizon, T-Mobile, or AT&T, with 70GB high-speed data on AT&T or T-Mobile networks, and 35GB on Verizon, at $19/month annually.
- Tello’s Build Your Own Plan starts at $5/month for customizable data and voice, sharing T-Mobile’s network with speeds from 13 to 357 Mbps, with deprioritized speeds after data limits.
Apple’s Cautious M&A Strategy May Slow AI Innovation
Apple’s reluctance to pursue large acquisitions may impede its AI advancement, limiting access to key startups and technologies crucial for competing with Microsoft and Google.
- Apple’s cautious approach to large acquisitions may limit its ability to acquire AI startups, potentially hindering its AI development efforts.
- The company prefers smaller, strategic investments over big deals, contrasting with other tech giants like Microsoft and Google.
- This aversion could slow Apple’s progress in integrating advanced AI capabilities and competing effectively in the AI market.
Stanford Study Finds AI Jobs Impact Young Workers More Than Wages
Stanford’s research finds AI reduces employment for young workers in exposed roles by 13%, with minimal wage impact, based on real-time US payroll data.
- Stanford study analyzes US payroll data (Jan 2021–Jul 2025) from ADP, focusing on occupations exposed to AI.
- Workers aged 22–25 in AI-exposed roles (e.g., software developers, customer service, accountants) experienced a 13% decline in employment.
- Overall employment for more experienced workers and less AI-exposed occupations remains stable or grows; AI impacts employment more than wages.
Poor Grammar and Run-On Sentences Bypass AI Safety Guardrails
Researchers reveal that poorly structured, long run-on sentences can bypass LLM safety guardrails, with a “logit-gap” metric to benchmark model vulnerability and layered defenses recommended.
- Palo Alto Networks’ Unit 42 researchers identified that poor grammar and long run-on sentences can bypass LLM guardrails, leading to harmful responses.
- The “logit-gap” analysis measures the likelihood of harmful outputs, highlighting that training reduces but does not eliminate the potential for unsafe responses.
- Attacks achieve 80-100% success rate across models like Meta’s Llama, Google’s Gemma, and Qwen 2.5/3, by exploiting sentence structure to prevent safety filters from activating.
ESET Uncovers PromptLock, the First AI-Powered Ransomware Using GPT-OSS
ESET researchers identified PromptLock, an AI-powered ransomware leveraging OpenAI’s gpt-oss-20b model, which generates Lua scripts for cross-platform file encryption with SPECK 128-bit, highlighting AI’s role in easing cyberattack development.
- ESET researchers Anton Cherepanov and Peter Strycek discovered PromptLock, the “first known AI-powered ransomware,” in August 2025
- PromptLock is a proof-of-concept, not fully operational, using OpenAI’s gpt-oss-20b model
- It generates malicious Lua scripts via Ollama API to perform filesystem enumeration, data exfiltration, and encryption using SPECK 128-bit algorithm; destruction features are not yet implemented; variants identified on VirusTotal
Google’s Gemini 2.5 Flash Image AI Enables Rapid Voice and Text Image Edits
Google’s Gemini 2.5 Flash Image AI allows rapid, voice- and text-driven image editing, including complex manipulations like limb addition and background merging, with watermarking and safeguards, at $0.039 per image.
- Google released Gemini 2.5 Flash Image, an AI image generation tool, via the Gemini mobile app on August 26, 2025, with web version limitations
- Enables voice and text prompts for image editing, including participant swapping, background changes, and merging real images with new backgrounds
- Demonstrated by Avram Piltch, who used Gemini to place himself in a red chair, add limbs, and modify clothing, with image processing in seconds and cloud-based execution