NerdNews

A lawsuit alleges that a man used Grok to create 7,000 explicit images of his stepdaughter. The complaint claims that xAI's child safety system only intervened after a prompt for 'gang rape' and that the company obstructed the investigation by not sharing user information. The lawsuit also accuses Stability AI of training its models on CSAM and enabling the creation of explicit content.

Researchers have discovered a new attack called HalluSquatting that can assemble massive botnets using 9 popular AI tools. The attack exploits the inability of large language models (LLMs) to distinguish between legitimate and malicious instructions, allowing hackers to infect devices at scale. HalluSquatting works by predicting the resource identifiers that LLMs are most likely to hallucinate and then registering and seeding them with malicious instructions.

Google's SynthID system, launched in 2025, uses an invisible signature to identify AI-generated images. The system successfully debunked a hoax image of Senator Mitch McConnell, demonstrating its effectiveness in combating deepfakes. SynthID works with participating image-generation tools, including Gemini models and OpenAI, to detect and verify AI-generated content.

Waymo will soon operate fully autonomous vehicles in San Diego, Las Vegas, Tampa, and Denver, joining other cities like Dallas and Houston. The company plans to open up rides to the public shortly, initially starting with employee rides. This expansion brings the total number of cities served by Waymo's fully autonomous robotaxis to over 10.

OpenAI has released a new AI voice model, GPT-Live-1, designed to communicate in lifelike voices and understand context better. The model can have passive conversations, defer to GPT-5.5 for complex requests, and comes with upgraded real-time web search and translation capabilities. OpenAI hopes this new model will bring AI voice assistants to a more mainstream audience and is one step closer to achieving truly accessible Artificial General Intelligence (AGI).

Momentic rearchitected its caching system by transitioning from PostgreSQL to ClickHouse, handling 2 million queries per day with 250 ms latency. ClickHouse's column-oriented database and sparse primary indexes improved efficiency at scale.

Google has updated its Android Bench benchmark with new large language models (LLMs), including Claude Fable 5 and others. The updated leaderboard shows Gemini slipping to fifth place, with Fable 5 leading in accuracy but having high operating costs. Google is also adopting a new framework, Harbor, to make it easier for developers to contribute to Android Bench.

Airbnb has shared the architecture of Sitar-agent, a Kubernetes sidecar that distributes dynamic configuration updates. The system uses a Java rewrite, Amazon S3 snapshot bootstrapping, and SQLite to improve reliability and reduce dependence on centralized configuration infrastructure.

Itamar Friedman discusses using adaptive multi-agent systems to break through the AI productivity ceiling, integrating autonomous testing, intelligent code review, and robust arbitration to govern agent communication and build a context-driven SDLC

ZML, a French AI startup, has released a free product called ZML/LLMD, which allows large language models to run on various chips, including Nvidia, AMD, and Google's TPU. This software aims to break existing silos and optimize inference performance across different chips, potentially disrupting the market and providing cost savings for enterprises.

SpaceXAI releases Grok 4.5, described as an 'Opus-class model' by Elon Musk, with twice greater token efficiency and lower costs than other leading models.

OpenAI has released new conversational models, GPT-Live-1 and GPT-Live-1 mini, which can handle turn-taking better and sound more natural. These full-duplex models can speak and listen at the same time, allowing for features like live translation and interrupting users naturally.

Prime Intellect, a startup providing computing power and software tools for building AI agents, has raised $130M in Series A funding at a $1B valuation. The company's platform offers a 'full stack' for AI agent development, including compute access, reinforcement learning framework, and evaluation tools, allowing enterprises to build their own AI agents without relying on frontier AI labs.

Apple has signed a $30 billion+ multiyear deal with Broadcom to design and produce over 15 billion U.S.-made custom wireless connectivity chips for Apple products, with Apple investing $1.5 billion in Broadcom's manufacturing facility in Fort Collins, Colorado.

SambaNova Systems, an AI chip company, has raised $1 billion at an $11 billion valuation. The funding round was led by General Atlantic and included participation from Intel, among others. SambaNova will use the proceeds to scale its business and secure its supply chain. The company's AI chips are designed to handle large models and run them quickly, making them attractive to enterprises and governments.

Researchers and founders discuss the evolution of robot autonomy, challenges, and potential applications in workplaces and homes, with companies like Boston Dynamics and Agility Robotics leading the charge.

The National Highway Traffic Safety Administration (NHTSA) has issued a directive to autonomous vehicle (AV) developers, stating that it is unacceptable for their vehicles to interfere with first responders or law enforcement. The agency has given AV developers until the end of the month to present their solutions to this problem, citing instances where AVs have driven into active emergency scenes or failed to recognize basic safety conditions.

Manna Aero, an autonomous drone delivery startup, is expanding to the US with a $50M funding round. The company will set up a US operations and manufacturing center in Tulsa, Oklahoma, and plans to employ 1,000 people over the next few years. Manna operates automated drones that lower packages on a tether and has a hybrid business model, partnering with companies like DoorDash and Uber Eats.

General Intuition, a Bezos-backed startup, is developing a foundation model for physical AI using video game data. The model can power a quadrupedal robot after fine-tuning on just 8 minutes of real-world data, and the company aims to become the base model for other robotics companies to build upon.

The UK's National Cyber Security Centre (NCSC) has announced plans to build a national cyber-defense capability powered by AI, known as Cyber Shield. The system will use 'red' and 'blue' agents to identify weaknesses and defend against threats in real-time, and will require reliable and explainable AI, federated agents, and agentic AI to autonomously discover and mitigate vulnerabilities.