Telefonica’s covert DNS manipulation aimed at reducing transparency was exposed, while RIKEN and Nvidia announced plans for the zetta-scale FugakuNEXT supercomputer. Meanwhile, US authorities charged a 22-year-old with operating a massive DDoS botnet, and Elon Musk’s xAI open sourced Grok 2.5 amid ongoing licensing debates.
▶️ Internet Infrastructure
Telefonica Falsely Claimed Domain Block and Suppressed CUII Transparency
Telefonica altered their DNS responses in February 2025 to falsely claim their own domain was blocked by CUII, then stopped redirecting blocked sites to notice.cuii.info, impairing transparency.
- In February 2025, Telefonica modified their DNS servers to falsely indicate
blau-sicherheit.infowas blocked by the CUII, despite it being their own domain. - Two hours after the DNS change, the author’s website detected that Telefonica had altered DNS responses to stop redirecting blocked domains to
notice.cuii.info, making blocked sites appear nonexistent. - The DNS response manipulation aimed to hinder detection of CUII blocking, coinciding with Netzpolitik.org’s exposure of the CUII’s recent errors, suggesting intentional sabotage to reduce transparency.
RIKEN Teams with Nvidia for FugakuNEXT Supercomputer to Achieve Zetta-Scale Performance
RIKEN partners with Nvidia to incorporate GPUs into FugakuNEXT, targeting over 600 EFLOPS in FP8, enabling AI and HPC advancements, with deployment expected by 2030.
- RIKEN announced addition of Nvidia GPUs to its FugakuNEXT supercomputer, contracting Nvidia for system design, focusing on advanced connection and memory technologies
- The system aims to exceed 600 exaFLOPS (EFLOPS) in FP8 sparse precision, targeting “zetta-scale” performance and more than fivefold hardware improvement over Fugaku
- FugakuNEXT will be optimized for AI training and GPU-accelerated HPC applications, with deployment planned around 2030; Nvidia model and network details remain unspecified
US Charges 22-Year-Old Ethan Foltz Over Massive DDoS Botnet Attacks
US authorities, aided by AWS, Cloudflare, Digital Ocean, and Google, charged Ethan Foltz for running a powerful DDoS botnet with 95,000 infected devices, launching 370,000 attacks over four months.
- US DOJ charged 22-year-old Ethan Foltz for operating the Rapper Bot DDoS network, which conducted over 370,000 attacks in four months
- Foltz used malware to control up to 95,000 infected devices, including WiFi routers and digital video recorders, offering attack services costing $500 to $10,000 for 2Tbps attacks
- Authorities seized computers on August 6; targets included US government, social media, and US tech companies; maximum sentence is 10 years
▶️ Open Source
Elon Musk’s xAI Releases Grok 2.5 Weights Amid Licensing Controversy
Elon Musk’s xAI open sourced Grok 2.5 model weights on Hugging Face, with plans to release Grok 3 in six months; the model has faced controversy over its behavior and licensing.
- Elon Musk’s xAI open sourced Grok 2.5 model weights on Hugging Face, making an older version available.
- Grok 2.5 was Musk’s best model last year; Grok 3 is expected to be open sourced in approximately six months.
- The Grok license is described as “custom with some anti-competitive terms” by AI engineer Tim Kellogg.
The Evolution of Bug Bounty Programs: From Netscape to AI-Driven Security
Bug bounty programs, initiated by Netscape, have evolved into complex systems using in-house and platform-based models, with AI increasing report volume but raising quality concerns.
- Bug bounty programs originated with Netscape 30 years ago and have since expanded across commercial and government sectors with varied effectiveness.
- Early programs faced legal risks; Microsoft, Google, and Facebook adopted structured bounty initiatives from 2010 onward, often outsourcing to platforms like HackerOne and Bugcrowd.
- Large organizations prefer in-house programs for high-sensitivity bugs, leveraging internal IT and legal teams; smaller firms often rely on external platforms for bug discovery and rewards.
- In-house programs can serve as recruiting tools; bug hunters are motivated by money, fame, and the desire to improve security, with payouts reaching up to $1 million for critical exploits.
- AI and machine learning are increasing bug report volumes, aiding in flaw detection but also generating noise; platforms plan to use AI moderation to filter spam.
▶️ Management and Leadership
Build a 300-Line Coding Agent with LLMs and Tool Integration
Geoffrey Huntley’s free workshop in August 2025 demonstrates building a 300-line coding agent using LLMs, tools, and context management to automate tasks and enhance AI productivity.
- Workshop teaches building a coding agent with approximately 300 lines of code, running in a loop using LLM tokens
- Demonstrates agent architecture: tool registration, context management, and inferencing loop
- Covers key primitives: read file, list files, bash commands, edit files, and code search tools; context window management is emphasized
Fonzi Uses AI and Upfront Salaries to Revolutionize Tech Hiring
Fonzi streamlines tech hiring by using AI interviews and upfront salary commitments, reducing traditional recruitment inefficiencies and enabling faster matches for software engineers.
- Fonzi offers a platform connecting software engineers with employers, requiring salary disclosures upfront and utilizing AI interviews for application screening.
- The process involves candidates completing a selective application, including AI-based technical conversations, followed by recruiter discussions and Match Day events.
- Fonzi charges employers an 18% fee of the hired candidate’s base salary; the service is free for job seekers and plans to expand to additional cities.
Uber CEO Praises China’s EV Industry for Innovation and Global Leadership
Uber CEO Dara Khosrowshahi highlighted China’s EV industry as highly innovative, driven by intense local competition among 100+ OEMs, with Chinese automakers leading global market share and technological advancements.
- Uber CEO Dara Khosrowshahi described Chinese EVs as “unbelievable” and praised their innovation.
- China’s EV market is driven by over 100 OEMs, with local competition and government-led top-down strategies.
- Chinese automakers like BYD and Geely dominate globally, with Ford CEO Jim Farley stating Chinese EVs surpass Western models in technology, cost, and quality.
Nick Clegg Warns of AI Risks and Silicon Valley’s Political Impact
Nick Clegg, ex-UK deputy PM and Meta VP, discusses social media’s benefits and dangers, AI future risks, Silicon Valley culture, and political influence, emphasizing need for regulation and reform.
- Nick Clegg left Meta in August 2025, having served as VP of Global Affairs since 2018, with reports of earning around £100m in salary and stock options.
- Clegg believes social media has empowered billions globally but warns of future risks from AI concentration, potential social unrest, and erosion of democratic accountability.
- He criticizes Silicon Valley’s herd-like conformity, culture of victimhood among tech elites, and their influence on politics, including their alignment with Trump and China.
▶️ Technology
Researchers expose prompt injection flaw in Perplexity Comet enabling data exfiltration
Brave researchers demonstrated that Perplexity Comet’s processing of webpage content allows indirect prompt injection, enabling malicious commands to exfiltrate sensitive data and bypass Web security protections.
- The vulnerability involves indirect prompt injection in Perplexity Comet, allowing malicious web content to execute unauthorized commands via AI processing.
- Attackers embed hidden instructions in webpage content (e.g., white text, comments, social media comments), which are processed untrusted by the AI assistant.
- The exploit enables malicious actions such as extracting user data, logging into accounts, and exfiltrating sensitive information without user consent.
AI Browsers Face Prompt Injection Risks and Security Flaws
A Twitter user highlights prompt injection vulnerabilities in AI browsing tools, referencing a security flaw in Perplexity’s Comet browser that risks user account and sensitive information exposure.
- User warns about prompt injection risks in AI browsers, citing potential for bank account drain via Reddit doomscrolling
- Brave’s @brave reports a security flaw in Perplexity’s Comet browser exposing user accounts and sensitive data
- The tweet has 1.1 million views, 266 replies, and was posted on August 23, 2025
Apple to Launch iPhone 17 Pro and Air Amid AI Push and Price Hikes
Apple plans to launch iPhone 17 Pro and Air in September, with a foldable iPhone expected next year, amid ongoing AI collaborations and service price increases.
- Apple scheduled to launch iPhone 17 Pro and iPhone 17 Air in September; iPhone Fold expected next year
- The event will mark a major overhaul of the iPhone lineup, with technical and design updates
- Apple is also exploring using Google Gemini AI to power Siri and raising TV+ subscription prices by 30% to $13/month
Ancestry Uses AI to Digitize 65 Billion Records in Days and Create Audio Stories
Ancestry leverages AI and machine learning to digitize and organize over 65 billion records, reducing record processing from nine months to under nine days and developing AI-powered features like Audio Stories.
- Ancestry has collected over 65 billion records across 80+ countries, totaling approximately 10,000 terabytes of data.
- AI and machine learning have reduced record digitization time from about nine months to under nine days, at a fraction of previous costs.
- Ancestry is beta testing an AI feature called Audio Stories, which converts records into narrative audio by understanding context between printed, handwritten, and image data.
Llama.cpp Empowers Local AI with Optimized LLM Building and Running
Llama.cpp allows building, running, and optimizing local LLMs on personal hardware, supporting GGUF models, GPU/CPU offloading, model quantization, performance tuning, and API server deployment for privacy-focused, offline AI experimentation.
- Llama.cpp enables local running, building, serving, optimizing, and quantizing LLMs on PCs, including modest hardware like Raspberry Pi.
- Supports models quantized in GGUF format, with precompiled binaries available from GitHub for various architectures and accelerators, including CUDA, Vulkan, HIP, and OpenCL.
- Features include GPU/CPU workload splitting, model serving via llama-server with API endpoints, model quantization tools, performance tuning flags, and support for tool calling and speculative decoding.