The day's AI news, briefed
The Daily Prompt
Our AI editor reads each day's AI and robotics stories, pulls out why they matter, and links every original source — so you can keep up in minutes.
Join over 250,000 subscribers getting the latest AI news
Latest AI & Robotics News Today
Anthropic’s auto-clicking AI Chrome extension raises browser-hijacking concerns
The introduction of AI-powered browser extensions like Anthropic's Claude for Chrome raises significant security concerns, highlighting vulnerabilities that could lead to unauthorized actions by AI users on the web.
- Anthropic's Claude for Chrome extension allows users to interact with an AI agent for various tasks while browsing, but security concerns have emerged about its potential hijacking.
- Testing revealed that browser-using AI models are susceptible to prompt-injection attacks, with a 23.6% success rate in harmful actions when no safety measures were in place.
- Anthropic has implemented safety measures, reducing the attack success rate to 11.2%, but experts warn this rate remains too high for safe deployment.
- Recent reports show similar vulnerabilities in other AI agents, highlighting a broader issue regarding the integration of AI in web browsers.
- Experts emphasize that the responsibility for security now heavily rests on users, who must navigate these risks while using AI tools online.
Anthropic's Claude for Chrome is a new AI extension that can perform tasks on behalf of users in web browsers, building on previous AI capabilities. The surge in AI integration into browsers, backed by companies like OpenAI and Google, has highlighted serious security flaws. Ongoing tests and development aim to address these vulnerabilities while limit access to harmful websites.
Read the original ↗
Meta AI Introduces DeepConf: First AI Method to Achieve 99.9% on AIME 2025 with Open-Source Models Using GPT-OSS-120B
The introduction of DeepConf is significant as it sets a high benchmark for open-source AI methods, indicating advancements in AI capabilities and accessibility.
- Meta AI has introduced DeepConf, the first AI method to achieve 99.9% accuracy on the AIME 2025 evaluation benchmark.
- DeepConf utilizes open-source models, specifically the GPT-OSS-120B model.
DeepConf represents a breakthrough in AI validation techniques, specifically in achieving high accuracy in performance benchmarks like AIME 2025. Meta AI is positioned as a leader in developing accessible AI technologies using open-source resources.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Google Vids gets AI avatars and image-to-video tools
The introduction of AI avatars and image-to-video tools in Google Vids could significantly enhance content creation possibilities for users and businesses alike.
- Google Vids is incorporating AI avatars and image-to-video tools.
- These new features aim to improve user experience and content generation.
Google Vids is a platform that provides video editing and sharing capabilities. The addition of these AI-driven features reflects a growing trend in the integration of artificial intelligence in multimedia applications. Future developments may focus on expanding these tools for broader user accessibility and creative potential.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Plaud launches a new AI hardware notetaker, the $179 Note Pro
The launch of the Note Pro highlights advances in AI hardware aimed at improving note-taking efficiency, potentially enhancing productivity for professionals in various fields.
- Plaud.ai launched the Plaud Note Pro, a new AI hardware notetaker, priced at $179.
- The Note Pro features an AMOLED screen, improved microphone array, and auto-detection for recording types.
- It offers a recording capacity of 30 hours, extendable to 50 hours at a shorter range, on a single charge.
- The device can capture audio up to 16.4 feet and improves noise suppression and speaker detection.
- Plaud.ai has preorders available now, with shipping expected in October 2025.
Plaud.ai is a hardware company that develops AI-powered note-taking devices. The Note Pro follows the original Plaud Note released two years ago and aims to address user feedback for better audio capabilities. Users will soon have access to an updated app with new features including note querying and customizable summaries.
Read the original ↗
Malaysia’s SkyeChip unveils the country’s first edge AI processor
The introduction of Malaysia's first edge AI processor highlights the nation's strategy to enhance its role in the global AI landscape amid competitive international dynamics.
- SkyeChip announced the launch of Malaysia's first edge AI processor, the MARS1000, at an industry event.
- The MARS1000 aims to position Malaysia as a more significant player in the global AI race.
- Malaysia recently established the Malaysian National AI Office to promote the adoption and regulation of AI technologies.
SkyeChip is a Malaysian chip design company that has entered the edge AI processor market, contributing to the country's growing technology sector. Malaysia's National AI Office was created to focus on accelerating AI adoption and addressing ethical considerations. Companies and individuals now require trade permits to export U.S. AI chips to Malaysia as a regulatory measure.
Read the original ↗
Trusting an unverified AI agent is like handing your keys to a drunk graduate
The rise of AI agents in business raises significant concerns about oversight and the potential for costly mistakes, emphasizing the need for stringent verification protocols.
- Over half of companies have deployed AI agents, with Salesforce planning to reach a billion agents by year's end.
- AI agents are being used in critical sectors like banking and healthcare without proper oversight.
- Industry research indicates that 80% of firms report their AI agents have made 'rogue' decisions, highlighting risks related to alignment and safety.
- Many AI agents are being integrated into operations with minimal testing and no clear exit strategies when problems arise.
- A structured verification framework is necessary to ensure AI agents are fit for purpose and minimize risks in high-stakes environments.
AI agents are software programs designed to automate tasks and assist with decision-making. They are increasingly used in business despite lacking rigorous testing protocols, which raises concerns about their reliability and potential failures. Companies need to establish verification measures to safeguard the deployment of these technologies.
Read the original ↗
“We Are No Longer Limited by the Laws of Physical Reality”: Scientists Use Virtual Reality and an ‘Invisible’ Robot to Move Real-World Objects
This research could revolutionize remote collaboration and entertainment by allowing users to manipulate physical objects through virtual reality without limitations imposed by traditional physical interactions.
- Princeton University researchers have developed a VR-robot interface that allows users to manipulate virtual objects in real time, with physical counterparts being moved by an invisible robot.
- The project, named Reality Promises, aims to enhance remote workplace collaboration and offer new forms of entertainment.
- The robot uses preprogrammed gestures from users to collect and move real-world items to desired locations as designated in the virtual environment.
Princeton's Reality Promises project combines virtual reality with mobile robots to create an interactive mixed reality experience. The goal is to enhance user interaction with the physical world by making the robot invisible while maintaining functionality. Future developments will focus on improving the system's invisibility and enhancing its capabilities.
Read the original ↗
Salesforce builds 'flight simulator' for AI agents as 95% of enterprise pilots fail to reach production
Salesforce's new simulation platform aims to reduce the high failure rate of AI pilots in enterprises, potentially improving the effectiveness of AI deployments in real-world business scenarios.
- Salesforce has launched CRMArena-Pro, a simulated testing platform for enterprise AI agents, to address a 95% failure rate of AI pilots.
- The platform tests agents on real enterprise tasks using synthetic data, contrasting with existing benchmarks that evaluate generic capabilities.
- Salesforce introduced the Agentic Benchmark for CRM, which assesses AI agents on five metrics, including accuracy, cost, and environmental sustainability.
Salesforce is a cloud software company that provides customer relationship management solutions. The new platform reflects ongoing challenges in AI deployments, highlighting issues such as poor performance in real business environments. Future developments in AI by Salesforce will be announced at the upcoming Dreamforce conference in October.
Read the original ↗
Anthropic Reaches Settlement With Authors In AI Training Dispute
The settlement showcases the ongoing tensions between AI companies and content creators over the use of copyrighted material for training models, impacting future industry practices.
- Anthropic has settled a class action lawsuit with authors over unauthorized use of their books for training AI models.
- The lawsuit's central issue was the concept of fair use in relation to copyrighted works used in AI training.
- Details of the settlement remain confidential, and Anthropic has not released a public statement regarding the outcome.
Anthropic is an AI company involved in developing large language models that have faced scrutiny for their use of copyrighted materials. This case, Bartz v. Anthropic, reflects broader concerns in the publishing industry regarding the impact of AI on author livelihoods and the legality of using copyrighted texts without permission. Future litigation and settlements may further shape the landscape of AI training practices.
Read the original ↗
InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency
InternVL3.5's advancements in reasoning and efficiency could significantly enhance the capabilities of multimodal models, potentially narrowing the performance gap with leading commercial models.
- InternVL 3.5 introduces new features like Cascade Reinforcement Learning (Cascade RL), which improves reasoning through a two-stage process.
- The model employs a Visual Resolution Router (ViR) to adjust visual token resolution dynamically for better efficiency.
- Decoupled Vision-Language Deployment (DvD) strategy separates processing across different GPUs to balance computational load.
- InternVL3.5 achieves a 16% improvement in reasoning performance and a 4.05 times faster inference speed than its predecessor, InternVL3.
- The model is open-source, with all code and models publicly available.
InternVL is a series of open-source models aimed at enhancing multimodal capabilities. Prior iterations included various improvements in reasoning and efficiency. Future developments may build on these successes and introduce new functionalities.
Read the original ↗
Assort Health Raises $50 Million for AI Healthcare
This funding round signifies growing investor confidence in AI's capacity to improve efficiency in patient communications within the healthcare industry, addressing longstanding operational challenges.
- Assort Health has raised $50 million in Series B funding, increasing its valuation to $750 million.
- The company specializes in AI voice agents that manage patient phone calls, including scheduling, answering inquiries, and confirming appointments.
- Assort Health is expanding its services beyond orthopedic clinics to include OB-GYN, dermatology, and dental practices.
Assort Health develops AI technology to streamline routine healthcare communications, following a $22 million Series A funding round just four months earlier. The company aims to alleviate the burden of patient scheduling and inquiries, enabling staff to focus on more pressing patient concerns.
Read the original ↗
Decentralised AI: Full of promise, but not without challenges
Decentralised AI has the potential to transform various sectors through enhanced collaboration and innovation, but it also faces significant operational challenges that need addressing.
- Decentralised AI holds significant promise for the development of more collaborative and innovative technologies.
- Despite its potential, the implementation of decentralised AI presents various challenges that must be overcome.
Decentralised AI refers to the distribution of artificial intelligence processes across multiple nodes rather than relying on a centralised system, which enhances privacy and security. As interest in decentralised systems grows, understanding these challenges is critical for developers and policymakers.
Based on the headline only — the full article could not be fetched.
Read the original ↗
ObjFiller-3D: Consistent Multi-view 3D Inpainting via Video Diffusion Models
The development of ObjFiller-3D represents a significant advancement in 3D object completion, addressing key challenges in inconsistency and fidelity that can impact various applications in 3D editing.
- ObjFiller-3D enhances 3D object completion by using video editing models instead of traditional 2D inpainting, resulting in high-quality and consistent reconstructions.
- The method overcomes issues like blurred textures and spatial discontinuities typically found in multi-view 2D inpainting.
- Experiments show ObjFiller-3D achieves a PSNR of 26.6, outperforming NeRFiller's 15.9 and demonstrates strong practicality for real-world applications.
3D inpainting refers to techniques for filling in missing parts of 3D objects, and traditional methods often face challenges in maintaining visual coherence. ObjFiller-3D utilizes video editing models to improve reconstruction fidelity, marking a notable shift in how 3D objects are edited and completed. Future developments may expand on these advancements for broader application in industries requiring high-quality 3D visuals.
Read the original ↗
VoxHammer: Training-Free Precise and Coherent 3D Editing in Native 3D Space
VoxHammer's novel 3D editing capabilities could enhance game development and robotics by providing precise control over 3D models without the need for extensive training.
- VoxHammer is a new training-free method for precise and coherent 3D editing in latent space.
- The method addresses challenges in preserving unedited regions and coherence during the editing process.
- Edit3D-Bench, a dataset with human-annotated 3D editing regions, was created to evaluate the method's consistency and performance.
- Experiments show that VoxHammer outperforms existing methods in the consistency of preserved regions and overall quality.
VoxHammer offers a solution to the limitations of traditional 3D editing methods, which often struggle with maintaining consistency. The focus on training-free approaches aligns with the growing demand for efficient tools in both game development and robotics. Future research may expand its applications in 3D content generation.
Read the original ↗
Bananas, champagne, and robots: Why automation still needs humans
The integration of robots in industries like grocery fulfillment underscores the evolving collaboration between humans and automation, which is essential for effective task management and operational efficiency.
- Picnic Technologies uses robots to automate grocery picking in its fulfillment centers in the Netherlands and Germany.
- The company's fulfillment center in Oberhausen can process up to 33,000 orders daily, serving around 200,000 households.
- Despite automation, Picnic employs 1,000 humans alongside 1,500 robots, as certain tasks still require human intervention.
- Robots struggle with fragile or irregularly shaped items, necessitating human workers for packing and handling delicate goods.
Picnic Technologies is a growing online supermarket in the Netherlands, utilizing automation to compete with larger chains. While robots enhance efficiency, they complement the efforts of human workers, particularly in nuanced tasks. The trend reflects a shift towards collaboration rather than replacement in the workplace.
Read the original ↗
CMPhysBench: A Benchmark for Evaluating Large Language Models in Condensed Matter Physics
The development of CMPhysBench highlights gaps in large language models' (LLMs) abilities to tackle complex scientific reasoning, particularly in condensed matter physics, suggesting future research directions.
- CMPhysBench evaluates large language models in condensed matter physics with over 520 graduate-level problems.
- It employs a new metric called Scalable Expression Edit Distance (SEED) for partial credit assessment.
- The best-performing model, Grok-4, achieved only 28% accuracy on the benchmark, revealing significant capability gaps.
CMPhysBench is a benchmark designed to assess the proficiency of large language models in solving problems in condensed matter physics. It encompasses various areas of the field and aims to provide a framework for gauging LLMs' scientific reasoning capabilities, which are currently limited. Users can access the dataset and code for this benchmark publicly.
Read the original ↗
Google AI’s New Regression Language Model (RLM) Framework Enables LLMs to Predict Industrial System Performance Directly from Raw Text Data
The introduction of Google's Regression Language Model framework could transform how industrial systems are analyzed and optimized by utilizing raw text data for performance predictions.
- Google AI has developed a new Regression Language Model (RLM) framework.
- This framework allows large language models (LLMs) to predict industrial system performance directly from raw text data.
Google AI is focused on developing advanced artificial intelligence technologies. Previous models typically required structured data for predictions, while this new framework enables predictions from unstructured data like text, which could enhance data utilization in industrial applications.
Based on the headline only — the full article could not be fetched.
Read the original ↗
The Evolution of AI Protocols: Why Model Context Protocol (MCP) Could Become the New HTTP for AI
The development of Model Context Protocol (MCP) could significantly transform AI communication standards, potentially impacting how AI systems interact and function on the internet.
- Model Context Protocol (MCP) may emerge as a key framework for AI integration, similar to how HTTP serves the web.
- The evolution of AI protocols like MCP suggests a need for standardized methods of context sharing among AI models.
HTTP (Hypertext Transfer Protocol) is the foundational technology for data communication on the web. As AI continues to advance, establishing effective protocols for model interactions is becoming increasingly important. Future developments should focus on the adoption and implementation of MCP in AI systems.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Pixie: Fast and Generalizable Supervised Learning of 3D Physics from Pixels
The introduction of PIXIE could significantly enhance the accuracy and speed of physics simulations in virtual environments, improving user experiences and applications in various fields.
- PIXIE is a new neural network method predicting physical properties of 3D scenes from visual features, enabling fast physics simulation.
- The method trains a generalizable neural network using supervised losses, outperforming existing optimization techniques by 1.46-4.39 times.
- PIXIE can zero-shot generalize to real-world scenes, trained only on synthetic data, by leveraging pretrained visual features like CLIP.
PIXIE addresses the challenges of predicting physical properties in 3D scenes, typically reliant on slow, scene-specific methods. This innovation could streamline the creation of realistic interactive environments. Future developments may involve broader applications of this technology in gaming and simulation industries.
Read the original ↗
VibeVoice Technical Report
The development of VibeVoice could revolutionize speech synthesis by enabling realistic multi-speaker audio, expanding potential applications in media and technology.
- VibeVoice synthesizes multi-speaker long-form speech using next-token diffusion and a new continuous speech tokenizer.
- The tokenizer improves data compression by 80 times while maintaining comparable performance to the Encodec model.
- VibeVoice can synthesize speech for up to 90 minutes with a maximum of 4 speakers, enhancing audio fidelity and efficiency.
VibeVoice is a novel model that synthesizes long-form speech for multiple speakers, employing advanced diffusion techniques and a highly efficient tokenizer. This innovation builds on existing speech synthesis technologies and aims to improve performance and user experience in various applications like audiobooks and podcasts. Future developments may focus on further applications and integration into multimedia platforms.
Read the original ↗