NerdNews

Anthropic's Claude Mythos Preview AI model is a general-purpose model that has been found to be extremely effective at identifying vulnerabilities in security systems, posing a significant cybersecurity risk. The company is only sharing the model with a limited number of major tech companies and banks through a program called Project Glasswing, to help them shore up existing security vulnerabilities and prepare for potential hacking attempts.

Cloudflare has announced a new infrastructure designed to run large AI language models efficiently. The infrastructure separates input processing and output generation onto different optimized systems and uses a custom inference engine to manage GPUs more efficiently. This allows for faster responses and more efficient use of resources.

A severe Linux security flaw called 'Copy Fail' has been uncovered, allowing any user to gain administrator privileges. The exploit, identified as CVE-2026-31431, affects nearly every Linux distribution released since 2017 and was discovered using AI scanning tools. Some distributions, including Arch Linux and RedHat Fedora, have already released patches, but many others remain vulnerable.

The Pentagon has struck deals with OpenAI, Google, Microsoft, Amazon, Nvidia, and xAI to use their AI tools in classified settings, leaving out Anthropic due to supply-chain risk concerns. The agreements will allow for the 'lawful operational use' of their systems, establishing the US military as an AI-first fighting force.

The article discusses the use of Codex, an AI tool, in various applications such as knowledge work, product management, and medicine. It highlights the potential of Codex to increase productivity and efficiency in these fields.

The Five Eyes security alliance has issued guidance on the use of agentic AI, warning that it may behave unexpectedly and amplify existing organizational frailties. The agencies recommend slow and careful adoption, prioritizing resilience over productivity, and highlight the need for strong governance, explicit accountability, and human oversight.

OpenAI is releasing its new GPT-5.5-Cyber model to a limited group of 'cyber defenders', despite previously criticizing Anthropic for similar exclusivity. The model can spot flaws, pentest, and exploit bugs, but its restricted access raises concerns about gatekeeping in AI development.

A Harvard study found that AI models, specifically OpenAI's o1 and 4o models, offered more accurate diagnoses than human doctors in emergency room cases. The study compared the diagnoses of 76 patients and found that the o1 model performed nominally better than or on par with two attending physicians. The researchers emphasized that the findings show an urgent need for prospective trials to evaluate these technologies in real-world patient care settings.

The article discusses how to deploy and configure local AI models like Qwen3.6-27B for coding on your computer, and explores agent frameworks like Claude Code, Pi Coding Agent, and Cline. It provides a step-by-step guide on setting up a local model and choosing an agent framework, and discusses the pros and cons of using local models versus cloud-based services.

OpenAI has introduced AI-generated pets for its Codex app, an agentic tool that helps with coding. These pets serve as floating overlays that provide updates on Codex's activities and can be summoned or dismissed using the '/pet' command. Users can choose from eight built-in pets or generate their own using the '/hatch' command, which utilizes AI.

HPE has expanded its ProLiant edge network portfolio to address the needs of businesses extending AI and mission-critical workloads to harsh environments. The new portfolio includes the HPE ProLiant Compute EL2000 chassis and HPE ProLiant Compute EL220 and EL240 Gen12 servers, which are optimized for edge deployments and validated for harsh and remote environments. The servers feature scalability, reliable operation in extreme temperatures, and durability in environments with heavy vibration and electromagnetic interference.

T-Mobile and Starlink have partnered to offer a new business internet service called SuperBroadband, which combines 5G and satellite connectivity to provide virtually unbreakable connectivity. The service aims to solve common problems of inconsistent reliability, limited coverage, and complexity in business internet, and is designed to help businesses stay online through outages and disruptions.

Over a third of new podcasts are now AI-generated, with companies like Inception Point AI producing thousands of episodes weekly, raising questions about their audience and purpose.

KC Green, creator of the 'This is fine' comic, claims AI startup Artisan stole his art for an ad campaign without permission. Artisan has reached out to Green, but he plans to seek legal representation.

Fujitsu confirms mainframe business will end in 2035, to be replaced by AI supercomputers or quantum computers. The company will focus on AI-driven management and shift to a value-based earnings structure. Fujitsu is also in talks with Japan, the UK, and Australia on defense tech that can contribute to global stability.

April saw 29 crypto hacks, with $651 million in losses, making it the worst month on record. Notable hacks include Drift and Kelp DAO, which combined for $579 million in losses. The incidents have raised questions about the security and decentralization of the crypto industry.

A Chinese court has ruled that it's illegal for companies to lay off employees and replace them with AI. The ruling states that employers are prohibited from shifting operating costs to employees and cannot use technological change as a pretext for unilaterally reducing salaries and terminating contracts.

OpenAI's CFO, Sarah Friar, reportedly wants to delay the company's initial public offering (IPO) from 2026 to 2027 due to concerns about the company's spending and revenue. This comes after OpenAI missed its recent revenue targets and is under pressure to become a leading player in the AI industry.

California's DMV has issued new rules for autonomous vehicle testing and deployment, including requirements for data collection and sharing, and a process for law enforcement to cite AV companies for traffic violations. The rules also allow heavy-duty vehicles equipped with autonomous vehicle tech to test and deploy on public roads.

Microsoft has released its first major update after CEO Satya Nadella vowed to 'win back fans'. The update includes several fixes, including improvements to Windows Explorer, and performance tweaks, such as faster startup app launching and improved storage performance. The update also includes enhancements to Windows Hello, the Microsoft Store, and the taskbar system tray. Additionally, the maximum FAT32 format size limit has been increased from 32GB to 2TB.