NerdNews

OpenAI will insert invisible watermarks into AI-generated text in the EU to comply with the EU AI Act, making it easier to identify AI-generated content. The watermark will be optional globally but mandatory in the EU.

OpenAI is testing visual ads in ChatGPT, which will appear when generating images. The ads will be separate from the image and won't influence ChatGPT's answers. They won't appear for subscribers of ChatGPT's Plus, Pro, or Enterprise plans.

OpenAI is introducing visual display ads to ChatGPT, which will appear alongside image generation results. The ads will be clearly labeled and won't influence ChatGPT's answers. This move aims to support free and low-cost subscription tiers and provide a new way for advertisers to reach ChatGPT's 1.2 billion weekly users.

Safeworld, founded by Dr. Ding Zhao, Kyle Wong, and Simo Rachidi, emerges from stealth with a $12M seed round to tackle the safety challenge of GenAI robots. The company uses simulations with realistic human models to evaluate robotic control systems, aiming to provide a third-party validation for robot-makers and ensure safe deployment of GenAI robots in unstructured environments.

OpenAI will add a digital watermark to text and code generated in the EU to comply with new AI transparency rules. The company is using a proprietary system called textGrain to add an invisible statistical signal to the model's word choices, which can be detected later. The watermark will not be enabled by default, except for eligible ChatGPT and Codex text output in the EU.

A vulnerability in AI agents from Google and others has exposed a structural flaw in the Model Context Protocol (MCP), allowing attackers to spread malicious instructions between agents. The flaw, dubbed 'protocol pivoting' or 'prompt injection,' can lead to server-side request forgery and other security issues.

Akka used 65 open source projects to test a spec-driven workflow for AI-assisted software porting, measuring specification structure, context, model, and effort selection. The experiment showed improved lines of code and performance in 57 of the 65 ports, with a 143,333 times improvement for one project.

This article discusses the importance of platform engineering for production Large Language Models (LLMs). It highlights key takeaways such as controlling hallucination rates, enforcing schema at the platform boundary, and employing versioned prompts. The article also explores the architecture of an LLM platform, including retry orchestration, intent validation, and authentication and authorization. It emphasizes the need for observability, cost attribution, and evaluation of agent behavior, and provides a summary of six concrete decisions for building an LLM platform.

Reflection AI debuts Beam, a 501-billion-parameter open-weight AI model that matches the performance of leading Chinese models at lower compute costs. Beam is designed for enterprises, the public sector, and developers, and is positioned as a 'workhorse model' for building customized AI systems.

TikTok introduces an AI shopping assistant and one-click checkout, allowing users to discover and purchase products directly from the app. The assistant provides real-time shopping guidance and remembers user preferences, while the checkout feature enables seamless transactions.

Adobe has filed a patent for an AI tool that generates navigable 3D worlds from text descriptions, potentially revolutionizing the field of 3D design and content creation. The tool uses a two-step process, first creating a 2D representation of the scene and then converting it into 3D geometry. This technology could save companies time and money in creating 3D environments, and has the potential to impact the gaming, VR, and AR industries.

Norway is introducing a temporary ban on AI glasses in public places due to privacy concerns, and will set up an expert group to propose permanent national rules. This move is part of a growing trend of governments regulating wearable technology, including smart glasses from companies like Meta.

OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EU using its new textGrain technology, which slightly changes word choices to create a detectable pattern. The watermark is not visible when reading or copying text and can be detected by approved researchers and organizations.

OpenAI is expanding ads in ChatGPT with a new visual format, allowing advertisers to promote products through images while users generate images. The company will begin testing this new ad format in the US with a group of advertisers, ensuring ads are clearly labeled and separate from the generated image.

Hackers stole 8 million citizens' records from Denmark's Central Person Register, including names, addresses, and social security numbers. The breach, discovered on October 2, is thought to be the biggest in the country's history.

The US has arrested a tech CEO for allegedly smuggling over $300 million in Nvidia chips into China. The CEO, Greg Lui, is accused of using false paperwork to mask shipments of servers containing export-controlled Nvidia chips. The FBI alleges that Lui conspired with freight-forwarding firms to illegally divert chip shipments into China, which could pose national security risks.

A US District Judge ruled that a warrantless search of automated license plate reader systems, including Flock, violated the Fourth Amendment. The judge suppressed the evidence obtained after the search, citing 'indiscriminate mass surveillance'. This decision may impact the use of Flock and similar technologies by law enforcement agencies.

A data breach in Denmark's Central Population Register exposed personal info of 8.8 million people, including names, addresses, and unique IDs. The breach occurred in September and was discovered on October 2. A private company's access to the registry was misused, and police are investigating.

ClingSTUN is a Linux proxy backdoor that exploits known vulnerabilities in IoT devices, using legitimate STUN servers to maintain connectivity and allow remote command execution. It has been observed exploiting 24 vulnerabilities across various devices and can spread through command injection flaws.

Researchers discovered new Linux backdoors targeting telecom and network-edge appliances, disguising their traffic as email and impersonating legitimate services. The backdoors, including BPFDoor and AVERAT, use SMTP protocol and hide in plain sight, making them difficult to detect.