NerdNews

Google announced that attackers prompted its Gemini AI chatbot over 100,000 times to clone its knowledge. The technique, called distillation, involves training a new model on a previous model's outputs. Google considers this intellectual property theft, but the line between standard distillation and theft is blurry.

Apple has patched a decade-old zero-day vulnerability in iOS that could allow attackers to execute arbitrary code. The flaw, discovered by Google's Threat Analysis Group, affects Apple's dynamic linker and was exploited in the wild. The patch is part of iOS 26.3, which also includes fixes for other bugs, including flaws that grant root access and disclose sensitive user information.

Anthropic has raised $30 billion in a Series G funding round, valuing the company at $380 billion. The round was led by Singaporean wealth fund GIC and investment management firm Coatue. The funding will be used to continue building enterprise-grade products and models.

OpenAI has released a new version of its Codex tool, powered by a dedicated chip from Cerebras, which enables faster inference and real-time collaboration. The new tool, called GPT-5.3-Codex-Spark, is designed for rapid prototyping and is currently available in a research preview for ChatGPT Pro users.

Waymo has announced its sixth-generation autonomous driving system, which will be used in its new Ojai robotaxis. The system features updated sensors, including a high-resolution 17-megapixel imager, and is designed to be more affordable and adaptable across different types of vehicles. Waymo plans to scale production at its Phoenix factory to tens of thousands of vehicles per year, with the goal of expanding its service to 20 additional cities this year.

OpenAI's new GPT-5.3-Codex-Spark coding model runs on Cerebras' Wafer Scale Engine 3, delivering code at over 1,000 tokens per second, 15 times faster than its predecessor. This move marks a significant shift away from Nvidia hardware, with OpenAI partnering with Cerebras to achieve faster inference speeds.

Cloudflare now converts HTML to Markdown for AI crawlers, reducing token count by 80% and making it cheaper for AI services to harvest site content. This feature complements Cloudflare's Content Signals Policy, which allows publishers to specify content usage preferences.

UJET's CEO says AI will improve call center agents' abilities to solve customer issues, rather than replacing them. Gartner predicts that by 2030, the cost per resolution for generative AI in customer service will exceed the cost of many B2C offshore human agents.

OpenAI has introduced its first model, GPT-5.3-Codex-Spark, designed to run on Cerebras Systems' AI accelerators. The model provides a more interactive experience for users of OpenAI's Codex code assistant, generating responses at over 1,000 tokens per second. This is made possible by Cerebras' waferscale architecture, which features ultra-fast on-chip memory called SRAM. The model is currently available in preview to Codex Pro users and via API to select OpenAI partners.

OpenAI has scaled a single primary PostgreSQL instance to handle millions of queries per second for ChatGPT, serving 800 million users. This was achieved through optimizations such as scaling up instance size, refining query patterns, and scaling out with additional read replicas. The primary instance is supported by nearly 50 geo-distributed read replicas, with reads distributed across replicas to maintain low latency.

Waymo has launched its 6th-generation autonomous driving system, which features improved sensors and AI-powered technology to navigate through extreme weather conditions. The system is being used in vehicles made by Chinese automaker Zeekr, sparking controversy among US lawmakers. The new system reduces the number of cameras needed and improves low-light sensitivity, resolution, and dynamic range.

Over 30 malicious Chrome extensions, disguised as AI chatbots, have been found to steal users' API keys, emails, and other sensitive data. These extensions, which have been installed by at least 260,000 users, use the same underlying codebase and communicate with the same infrastructure, according to LayerX Security. The extensions impersonate popular chatbots like Claude, ChatGPT, and Gemini, and can extract data from websites, including Gmail, using iframes and content scripts.

Spotify's best developers haven't written code since December due to AI. The company uses an internal system called 'Honk' for remote, real-time code deployment using generative AI, specifically Claude Code. This system has sped up coding and deployment 'tremendously' and has allowed engineers to add new features and fix bugs quickly.

Aurora's self-driving trucks can now travel 1,000 miles nonstop, exceeding human drivers' capabilities. This breakthrough can cut transit times nearly in half, offering compelling economics to customers. Aurora plans to expand its driverless routes across the Sun Belt, with a fleet expected to grow to over 200 trucks by the end of the year.

The Pitt's AI drama mirrors real-life debates in hospitals. Two-thirds of physicians use AI, but many fear its errors and risks. AI tools can improve efficiencies, but also raise concerns about de-skilling and patient safety.

The latest viral AI apocalypse warning is overstated. While AI is improving rapidly, its impact on the economy will be gradual, not instantaneous. Several bottlenecks, including the need for heavy investment in data infrastructure and worker retraining, will slow its adoption.

Samsung and Micron have started shipping HBM4 memory, which is faster and denser than its predecessor. This memory is needed to power the next generation of AI acceleration hardware. Samsung's HBM4 delivers a consistent processing speed of 11.7 gigabits-per-second and can reach 3.3 terabytes-per-second in a single stack. The news is significant for Nvidia, which plans to release its Vera Rubin accelerators in the second quarter of 2026 and will use HBM4 memory.

Threat actors are using legitimate employee monitoring software to blend into corporate networks and attempt ransomware deployment. In two recent incidents, attackers used Net Monitor for Employees Professional and SimpleHelp to gain access to victim computers and deploy ransomware. The attackers were ultimately unsuccessful, but the incidents highlight the need for companies to be vigilant about the use of employee monitoring software and to take steps to prevent ransomware attacks.

Elon Musk unveiled a new vision for SpaceX and xAI, involving a moon base called Moonbase Alpha. The base would be used to manufacture space computers and launch them into deep space using a mass driver. Musk believes this could help achieve a new level of energy usage, as described by the Kardashev Scale, and potentially lead to the development of a solar system-scale supercomputer.

Didero, a startup founded by Tim Spencer, has raised $30 million in Series A funding to automate manufacturing procurement using agentic AI. The platform sits on top of a company's existing ERP, reading incoming communications and executing updates and tasks. Didero aims to streamline the procurement workflow, from sourcing to payment, for manufacturers and distributors.