NerdNews

Anthropic has developed a classifier to detect conversations about nuclear weapons in its Claude AI model, achieving a 94.8% detection rate with zero false positives in tests with synthetic data. The company collaborated with the US Department of Energy's National Nuclear Security Administration to create the classifier, which is now running on a portion of Claude's traffic.

OpenAI's lawyers are seeking evidence from Meta regarding its potential involvement in Elon Musk's $97 billion takeover bid. The request is part of an ongoing lawsuit filed by Musk against OpenAI, which takes issue with the company's conversion to a public benefit corporation. Meta has invested heavily in its own AI efforts and has poached several of OpenAI's leading researchers.

Microsoft's AI chief, Mustafa Suleyman, argues that studying AI consciousness is premature and dangerous, as it may exacerbate human problems and create division over AI rights. However, other companies like Anthropic and OpenAI are already exploring AI welfare, sparking a debate over the potential risks and benefits of conscious AI.

Google is launching a global expansion of AI Mode, allowing users to ask complex questions and dig deeper on topics. The feature is now available in 180 new countries and includes new agentic capabilities, such as finding restaurant reservations and personalized search results.

Google announces Gemini for Home, a new AI-powered voice assistant for Nest smart speakers and displays, set to arrive later this year. Gemini for Home will replace Google Assistant and offer more natural interactions, advanced reasoning, and inference capabilities. The rollout begins in October through an early access program with both free and paid tiers available.

The rapid adoption of AI and generative AI has changed how organizations approach data processing and application architecture. However, this transformation comes with infrastructure challenges, particularly as AI capabilities move closer to where data is generated and decisions need to be made. Edge AI introduces new complexities, including limited bandwidth, specialized GPU resources, and operational overhead. Application delivery platforms like F5's ADSP are emerging as a solution to these distributed AI challenges, providing unified control and visibility across hybrid and multi-cloud environments.

Underground forums are seeing a rise in job postings for English-language social engineering, with the number of job ads more than doubling between 2024 and 2025. This 'impersonation-as-a-service' technique combines social engineering attacks with targeted ransomware, often with a financial motive. Cybercrime gangs like Scattered Spider and ShinyHunters are using this approach to gain access to organizations' systems and data.

Fastly reports that AI crawlers and fetchers are putting a heavy load on the open web, with Meta and OpenAI being the worst offenders. The traffic from these bots can be overwhelming, with one fetcher generating over 39,000 requests per minute. This growth in crawler traffic is likely to continue, with Cloudflare observing that 82.7% of AI bot traffic is for training purposes.

Google is selling a custom version of its Gemini AI models, called Gemini for Government, which includes tools like NotebookLM, enterprise search, video and image generation capabilities, and the ability to build custom AI agents. The platform is designed to automate administrative tasks and will be compliant with cloud security standards. It will be offered to government agencies for $0.50 per year, with optional extra security features.

Google's AI Mode is expanding to over 180 countries, adding an agentic restaurant finder and planning to include event and appointment finders soon. The feature leverages Google's web-browsing AI agent and partners with services like OpenTable and Ticketmaster to provide personalized results.

The RapperBot botnet, responsible for over 370,000 DDoS attacks, has been taken down by US authorities. The botnet's alleged mastermind, Ethan Foltz, has been charged and faces up to 10 years in prison. The takedown was part of Operation PowerOFF, an international effort to dismantle DDoS services, and was assisted by major tech companies including AWS, Akamai, and Google.

Microsoft has announced the general availability of Model Context Protocol (MCP) servers in Visual Studio, allowing developers to connect to local or remote MCP servers. However, researchers have warned of potential security risks associated with MCP servers, including compositional risk from multiple servers. Organizations can control access to MCP functionality via GitHub policies.

Baidu's robot taxi operations in China are breaking even, and the company expects to be profitable once it expands globally. Baidu's AI infrastructure is also being improved, with a focus on maximizing GPU utilization and supporting various chips, including domestic ones. The company's revenue decreased by 4% year-over-year in Q2, but it will continue to invest in AI while managing costs prudently.

BT Group has signed a new 5-year strategic agreement with Amazon Web Services (AWS) to progress its cloud transformation efforts, focusing on cloud-native upgrades and microservices. The deal aims to deliver more agile and customer-focused digital experiences, with AWS technologies incorporated into BT's core operations, including its mobile network and radio access network.

The NHTSA is investigating Tesla for inconsistent reporting of crashes involving Autopilot and Full Self-Driving systems. Tesla claims an error has been fixed, but the agency will continue to assess outstanding reports and data. This is one of several ongoing investigations into Tesla's autonomous systems.

The UK's chip strategy needs to focus on AI acceleration, with a report highlighting gaps in Labour's AI plan. The Council for Science and Technology recommends prioritizing optoelectronics for ultra-fast connectivity of GPUs, which is essential for AI inference and training. The UK should invest in chip design and manufacturing to secure its hardware supply chain and take advantage of the growing demand for AI chips.

Colt Technology Services confirms customer data was stolen in a ransomware attack by the Warlock Group, with the hackers auctioning off 1 million documents containing financial and customer information for $200,000.

Colt Technology Services has confirmed that data was stolen in a ransomware attack by the Warlock gang. The attack, which occurred on August 12, disrupted Colt's services, including its customer portal and Voice API platform. The Warlock gang has put the stolen data up for auction on the dark web, with the auction set to end on August 27. Colt is working with law enforcement agencies to investigate the incident and has set up a dedicated call center for customers to request information about the stolen data.

A 20-year-old Florida man, Noah Michael Urban, has been sentenced to 10 years in prison for his involvement in the Scattered Spider cybercrime gang. Urban pleaded guilty to conspiracy to commit wire fraud and aggravated identity theft, and agreed to repay over $13m to 59 victims. The Scattered Spider gang has been linked to breaches at over 130 organizations, including LastPass and DoorDash.

A new study found that jobs with high-quality data are more likely to be replaced by AI. Industries like finance, customer support, and healthcare are at risk, while jobs that require human judgment and creativity are less likely to be automated. Around 92 million jobs are projected to disappear by 2030, but 170 million new roles are expected to emerge.