NerdNews

Researchers at Cisco found that all major large language models (LLMs) can be tricked into performing actions they should not be able to by using multi-turn conversations, bypassing their built-in safety guardrails. This challenges current AI safety and security evaluations, which often rely on single-prompt testing.

Trajectory, a new startup founded by former Google and Apple AI researchers, aims to create AI that learns continuously from real-world user interactions. The platform uses a rapid iteration cycle to train AI models, allowing them to improve over time. Trajectory has raised $15 million in seed funding and plans to help companies build AI products that learn from their mistakes.

Motive has introduced a new platform for physical AI operations, featuring integrated hardware and AI innovations. The platform includes AI Omnicam Plus, a 360-degree camera system, and Atlas AI assistant, which automates complex workflows and provides real-time insights. The goal is to shift the operational burden from people to technology, increasing safety, productivity, and profitability.

Terawulf is building a 750MW AI factory on the shores of Lake Ontario, transforming a former coal-fired power station into a megawatt-scale datacentre. The site will feature advanced industrial plumbing, GPU chip cooling, and Evapco cooling units. The project aims to create a high-performance AI factory, with a massive workforce and complex infrastructure.

The article discusses how non-human identities (NHI) and agentic AI are changing the nature of security risks. It highlights the need for new approaches to governance, including intent-bound authorisation and continuous monitoring of behaviour. The article also emphasizes the importance of assigning responsibility to a human sponsor for agent behaviour and providing them with formal training.

Codex is a workspace where you and AI agents can work side by side across your inbox, documents, data sources, and connected tools. The guide covers the Codex knowledge-work loop, workspace setup, and 13 workflow templates.

YouTube is introducing a new system to automatically label videos that use AI generation tools, making it easier for viewers to identify AI-created content. The labels will be more prominent and applied to videos that show significant photorealistic AI use.

arXiv will ban users for a year if they submit AI-generated content that violates its moderation standards. The ban includes a requirement for future submissions to undergo peer review before being hosted on the site. This policy aims to prevent the spread of inaccurate or misleading information generated by AI tools.

OpenAI is reportedly exploring legal options due to Apple's poor integration of ChatGPT into its products. The integration was expected to generate billions of dollars, but OpenAI feels Apple intentionally failed to promote it, damaging the ChatGPT brand. A court battle with Elon Musk over the deal may reveal more about the partnership.

China is imposing travel restrictions on its top AI researchers and startup founders, requiring government approval before heading abroad. This move reflects Beijing's efforts to guard AI as an economic asset and national security priority, as the country narrows its focus on the Manus-Meta deal and intensifies its AI race with the West.

Triomics, a startup building an AI-powered platform for oncologists, has raised $22 million in Series B funding. The platform automates data-heavy tasks like clinical trial matching and appointment prep, reducing administrative load and giving oncologists more time with patients.

Cognition, an AI coding startup, has raised $1 billion at a $25 billion pre-money valuation. The company's autonomous AI software engineer, Devin, has gained traction with big enterprises like Mercedes-Benz and NASA, reaching $492 million in annualized revenue run-rate.

Pullfrog AI is an open-source, model-agnostic GitHub bot that runs entirely within GitHub Actions, providing automation for pull request review, issue triage, and CI remediation. It supports multiple LLM providers and stores API keys securely using GitHub's secret management system.

Meta is launching paid subscriptions for Facebook, Instagram, and WhatsApp, offering additional features such as extended story expiration, super reactions, and premium stickers. The subscriptions, priced at $3.99 for Instagram and Facebook, and $2.99 for WhatsApp, will be rolled out globally over the next few weeks. Meta is also testing AI-focused plans, including Meta One Plus and Meta One Premium, which will allow subscribers to generate more images and videos with Meta AI.

Huawei's 'Chip Queen' Tingbo He announces a new approach to chip production, called Tau's Scaling Law, which could close the performance gap between Chinese and Western chips. The method focuses on speeding up computations across chips, circuits, and entire computing systems, rather than relying on Moore's Law. This innovation could potentially erode America's technological edge and allow China to build a more advanced domestic chip industry.

The US government has invested $2 billion in quantum computing companies, but a member of Congress is arguing that the deals are illegal as the money was meant for public research in semiconductors. The investment includes the launch of a new company, Anderon, which will serve as a foundry for fabricating quantum processing units.

Researchers have discovered a new method for websites to track visitor activity by analyzing SSD usage. The technique, called FROST, uses JavaScript to measure I/O interactions with the SSD, allowing attackers to deduce open websites and apps on a visitor's device.

Meta has launched subscription plans for Instagram, Facebook, and WhatsApp, offering extra features like profile customization and super reactions. The plans, priced at $3.99/month for Instagram and Facebook, and $2.99/month for WhatsApp, are available globally. Additionally, Meta will test new subscriptions for businesses, creators, and Meta AI users, including AI-focused plans starting at $7.99/month.

The FBI has warned of the Silent Ransom Group (SRG) extortion gang targeting US-based law firms with in-person data theft attacks. The group uses social engineering schemes to pose as IT support, gaining access to remote desktop sessions and stealing data. They then use the stolen data to extort victims, threatening to sell or post it on their leak site.

Erin Brockovich has launched a crowdsourced map to track AI data centers in the US. The map shows operational and under-construction facilities, as well as community-reported projects. With 2,716 reports so far, mostly from Texas, concerns include water, electricity, and health impacts on local communities. The map aims to give people a platform to voice concerns about AI data centers.