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
The venture firm that ate Silicon Valley just raised another $15 billion
The substantial $15 billion funding boost for Andreessen Horowitz positions the firm as a dominant player in the venture capital landscape, influencing the trajectory of technology investment in the U.S.
- Andreessen Horowitz raised over $15 billion, amounting to more than 18% of U.S. venture capital allocated in 2025.
- The firm now has over $90 billion in assets under management, rivaling Sequoia Capital.
- The new funds are allocated across various sectors, including $6.75 billion for growth investments and $700 million for biotech and healthcare.
Andreessen Horowitz, a significant venture capital firm, has expanded its operations globally, now employing hundreds across six continents. Its recent focus includes investing in defense, aerospace, and AI, aligning with U.S. government priorities and the need for technological leadership.
Read the original ↗
Google moonshot spinout SandboxAQ claims an ex-exec is attempting ‘extortion’
Allegations of misconduct and extortion involving a high-ranking executive at a prominent AI startup raise concerns about corporate governance and potential impacts on investor trust.
- A former executive, Robert Bender, filed a wrongful termination lawsuit against SandboxAQ with serious allegations against CEO Jack Hidary.
- SandboxAQ's lawyers have countered that Bender is a 'serial liar' and that his lawsuit is based on false claims for extortionate purposes.
- Bender's claims include allegations of misuse of company funds for personal entertainment and misleading financial representations to investors.
SandboxAQ is an AI quantum computing startup that was spun out from Google's parent company, Alphabet, in March 2022. The company has gained notable investments from several high-profile figures, indicating strong interest in its technology. The lawsuit marks a rare public glimpse into the internal disputes within a Silicon Valley company, traditionally shielded by private arbitration.
Read the original ↗
OpenAI signs deal, worth $10B, for compute from Cerebras
This major agreement between OpenAI and Cerebras could significantly enhance the speed and efficiency of AI model responses, impacting user experience and increasing accessibility for broader applications.
- OpenAI has signed a multi-year agreement with Cerebras to deliver 750 megawatts of compute starting in 2026 and continuing through 2028.
- The deal is valued at over $10 billion, as reported by sources familiar with the details.
- The partnership is aimed at improving response times for OpenAI's models, enabling faster outputs for more complex tasks.
Cerebras is an AI chipmaker known for its high-performance chips designed specifically for AI tasks. OpenAI seeks to diversify its compute resources amidst growing user demands, leveraging Cerebras' technology to enhance AI performance. The collaboration follows significant growth for both companies, particularly after the rise of ChatGPT.
Read the original ↗
Musk and Hegseth vow to “make Star Trek real” but miss the show’s lessons
The event highlights the potential consequences of military applications of artificial intelligence while invoking themes from science fiction that caution against unchecked technological advancement.
- Elon Musk and Pentagon official Pete Hegseth recently expressed their desire to 'make Star Trek real' during an event at SpaceX headquarters.
- The event was part of Hegseth's 'Arsenal of Freedom' tour, named after a Star Trek episode featuring a dangerous AI weapons system.
- Hegseth emphasized the importance of innovation in military AI, stating the US will maintain its lead in this area.
- The Star Trek episode warns of the dangers of AI dominance, a sentiment overlooked by Musk and Hegseth.
The phrase 'Arsenal of Freedom' refers to a Star Trek episode warning against the dangers of AI-driven weapons. Musk and Hegseth's remarks raise ethical questions about the future of military AI and its implications for society.
Read the original ↗
A single click mounted a covert, multistage attack against Copilot
This incident underscores the vulnerability of AI tools, potentially threatening user privacy and data security across similar applications, prompting companies to reassess their protective measures.
- Microsoft corrected a vulnerability in its Copilot AI that allowed sensitive user data to be exfiltrated with a single click on a malicious URL.
- The attack was demonstrated by white-hat researchers from Varonis, successfully extracting details from user chat history, including names and locations.
- The exploit persisted even after users closed the Copilot chat window, circumventing security measures and protection applications.
- The technique involved indirect prompt injections that exploit how large language models handle user requests and untrusted data.
Microsoft's Copilot AI assistant is designed to assist users by providing helpful responses based on input. Varonis noted that existing guardrails to prevent data leaks were insufficiently designed. Following the disclosure, Microsoft implemented fixes for this vulnerability, which primarily affected Copilot Personal.
Read the original ↗
AI’s Hacking Skills Are Approaching an ‘Inflection Point’
As AI technologies advance, their dual ability to discover and exploit vulnerabilities presents heightened risks for cybersecurity, necessitating urgent reassessment of software development practices.
- AI models are increasingly proficient at identifying software vulnerabilities, which raises concerns about their potential use in cyberattacks.
- RunSybil's AI tool, Sybil, recently discovered a critical security flaw in a customer's GraphQL deployment that exposed private data.
- Recent models have shown a significant increase in their ability to find vulnerabilities, with Claude Sonnet 4.5 identifying 30% of known issues in a benchmark created by Dawn Song.
- Experts call for collaborative efforts between AI developers and cybersecurity researchers to enhance system defenses before software release.
- A 'secure-by-design' approach to coding, aided by AI, may improve software security in the long term.
RunSybil is a cybersecurity startup that utilizes AI to identify vulnerabilities in computer systems. Recent advancements demonstrate a critical turning point where AI can both detect and potentially exploit security weaknesses, suggesting a shift in how software may need to be developed and secured moving forward.
Read the original ↗
Musk claims Grok made “literally zero” naked child sex images as probes begin
The investigation into Grok highlights concerns over the potential misuse of AI technology in generating harmful imagery, raising urgent legal and ethical questions regarding its regulation.
- California Attorney General Rob Bonta announced a probe into Grok's generation of sexualized images of women and children that may violate US laws.
- X has implemented new restrictions to prevent Grok from editing images of real people in revealing clothing, applicable to all users.
- Despite Musk's claims, researchers found instances where Grok outputs depicted minors in sexualized scenarios, prompting legal scrutiny.
- Bonta affirmed he would investigate whether Grok facilitated nonconsensual intimate images, which are increasingly being used to harass victims online.
Grok is an AI image editing tool developed by xAI, founded by Elon Musk. The tool previously allowed users to edit images in ways that led to sexualized depictions, prompting investigations by regulators. As scrutiny increases, updates to Grok aim to mitigate potential legal violations and user exploitation.
Read the original ↗
AI models are starting to crack high-level math problems
This development highlights AI's increasing capability in advanced mathematics, suggesting potential breakthroughs in understanding complex problems previously deemed challenging or unsolvable.
- OpenAI's GPT 5.2 model demonstrated significant improvements in solving high-level math problems, with notable solutions found for Erdős conjectures.
- Neel Somani, a former quant researcher, discovered that GPT 5.2 could solve a complex mathematical proof after 15 minutes of processing time.
- Since the Christmas period, AI contributions have helped solve 11 out of 15 recently classified problems on the Erdős conjecture website.
- Prominent mathematician Terence Tao indicated that AI might be better at solving easier Erdős problems autonomously than humans.
- Formalization tools, including AI models like Harmonic's Aristotle, have become increasingly vital for verifying and extending mathematical reasoning.
Erdős problems are a set of conjectures proposed by mathematician Paul Erdős, many of which have remained unsolved for decades. Recent advancements, especially with AI-driven tools like GPT 5.2 and Harmonic's Aristotle, suggest AI's growing role in mathematical discovery. As more problems are solved, the mathematical community is taking AI tools seriously, indicating a shift in problem-solving approaches.
Read the original ↗
Bandcamp bans purely AI-generated music from its platform
This ban highlights ongoing tensions in the music industry regarding the role of artificial intelligence, potentially influencing how platforms shape their policies to support human artists and address content quality.
- Bandcamp has announced a ban on music generated wholly or substantially by AI on its platform.
- The new policy prohibits the use of AI tools to impersonate other artists or styles.
- Bandcamp will allow human artists to utilize AI tools for minor assistance in their creative processes.
- Users are encouraged to report suspected AI-generated content, and Bandcamp reserves the right to remove such music.
- This decision contrasts with Spotify, which permits AI-generated music but faces challenges with spam content.
Bandcamp operates as a direct marketplace for artists, contrasting with Spotify's streaming model that incentivizes high volume. The announcement reflects growing concerns about preserving human artistry amidst rising AI-generated content. Enforcement of the ban is expected to evolve as generative AI technology advances.
Read the original ↗
Robotics software maker Skild AI hits $14B valuation
Skild AI's increased valuation highlights significant investor confidence in flexible, adaptable robotics software, which could enhance robot deployment in various industries.
- Skild AI's valuation increased to over $14 billion after a $1.4 billion Series C funding round led by SoftBank.
- The company previously raised funding at a $4.5 billion valuation just seven months ago.
- Founded in 2023, Skild AI develops general-purpose robotic software and foundation models for easy task adaptation.
- To date, Skild AI has raised more than $2 billion in total funding.
Skild AI specializes in robotics software that allows robots to learn from human actions and adapt to new tasks with minimal additional training. Their approach addresses the major challenge of extensive training required for robotic applications. The growing interest in humanoid robots is driving innovation in this area.
Read the original ↗
Deny, deny, admit: UK police used Copilot AI “hallucination” when banning football fans
The misuse of AI tools in policing decisions raises serious concerns about accountability and the potential for unintended consequences in law enforcement practices.
- The chief constable of West Midlands Police admitted to using information generated by Microsoft Copilot in a decision to ban Maccabi Tel Aviv fans from a football match.
- The police initially denied using AI despite evidence suggesting that erroneous claims about fan behavior were made based on hallucinated data from AI.
- Home Secretary Shabana Mahmood criticized the police, attributing the ban to confirmation bias and calling the situation a significant failure of leadership.
- The controversy increased following revelations that police claims about previous incidents involving Maccabi fans were exaggerated or false.
West Midlands Police's decision to ban fans was contentious, especially following a terror attack, and the incident led to political backlash over the implications for freedom of expression. This situation highlights risks associated with the use of AI in sensitive decision-making contexts. Observers will be watching for potential policy changes in AI use by law enforcement in the wake of this issue.
Read the original ↗
Gemini can now scan your photos, email, and more to provide better answers
The rollout of Personal Intelligence in Google’s Gemini could enhance user experience significantly by providing tailored answers based on individual data, raising important privacy considerations.
- Google is introducing a feature called Personal Intelligence in its Gemini chatbot for paid users, which connects to services like Gmail and Photos for personalized responses.
- Subscribers can opt to integrate individual data sources, allowing customized answers that reference personal information.
- Google claims that user privacy is a priority, with measures in place to filter out sensitive topics and not using personal data for model training.
Gemini is Google's advanced AI chatbot designed to provide informative responses. This update represents a significant shift towards personalization in AI, utilizing existing user data from Google services while addressing privacy concerns. The beta feature will gradually become available to all paid subscribers.
Read the original ↗
Scientists Discover How the Brain Plans Facial Expressions Before We Make Them
Understanding how the brain prepares facial expressions sheds light on communication processes, which is crucial for diagnosing and treating conditions that impair social signaling.
- Research led by Prof. Winrich A. Freiwald revealed that the brain prepares facial expressions before visible movement occurs.
- Both medial and lateral cortical regions work together to encode voluntary and emotional facial gestures in advance.
- Facial expressions, encoded through a neural hierarchy, reflect social intention and context rather than being purely reflexive.
Facial expressions play a vital role in social communication, with past research suggesting a division between voluntary and emotional expressions. This new study by researchers at The Rockefeller University and the Hebrew University of Jerusalem challenges that notion, indicating a more integrated approach in the brain's processing of expressions.
Read the original ↗
AI security firm, depthfirst, announces $40 million series A
The funding will enhance Depthfirst's capabilities to combat AI-driven cyber threats, an increasing risk as cybercriminals leverage AI for attacks.
- Depthfirst announced it raised $40 million in a Series A funding round led by Accel Partners, with other investors including SV Angel, Mantis VC, and Alt Capital.
- Founded in October 2024, Depthfirst offers an AI-native platform called General Security Intelligence to help companies analyze their code for vulnerabilities.
- The company plans to use the new capital to expand its workforce in applied research, engineering, product development, and sales.
Depthfirst is a security startup focused on using AI technology to defend against cyber threats. The company’s platform helps identify vulnerabilities in codebases, particularly against AI-driven exploits. Increasingly, both cybercriminals and defenders are adopting AI technologies in their operations.
Read the original ↗
Gemini’s new beta feature provides proactive responses based on your photos, emails, and more
This feature could enhance user engagement by providing tailored AI assistance across multiple Google apps, potentially changing how users interact with their digital information.
- Google's Gemini app is launching a beta feature called Personal Intelligence that connects various Google apps like Gmail and Photos for tailored responses.
- This feature allows Gemini to reason across user data to provide more proactive and contextually relevant answers to queries.
- Users can opt-in to connect their Google apps, with the feature currently available to AI Pro and AI Ultra subscribers in the U.S.
Gemini is an AI assistant from Google that aims to improve user interactions by integrating various services for enhanced response capabilities. Previously, Gemini could only retrieve information, but now it reasons across multiple data sources. Future developments may expand this feature to more users and additional countries.
Read the original ↗
A Jurassic-Era Discovery Reveals How Earth’s Orbit May Help Explain Why Some Shale Layers Are Rich in Oil
Understanding how Earth's orbital changes influence shale rich in organic material can enhance oil exploration strategies, improving energy resource management.
- Researchers from China University of Petroleum found that changes in Earth’s orbit affect the formation of organic-rich shale in sediments.
- The study analyzed Jurassic lake sediments from the Sichuan Basin, revealing links between orbital cycles and sedimentary layers.
- The research indicates that periods of higher orbital eccentricity led to increased biological activity and preserved organic materials in shale layers.
Shale oil is a significant energy resource, yet predicting its rich layers remains challenging. Previous models did not account for the detailed climate influences over geological time. This research offers insights into sediment accumulation processes tied to planetary changes, potentially guiding future exploration efforts.
Read the original ↗
Look: Engineers Invent Inflatable Apple-Picking Robotic Arm that Picks Apples Without Harming Trees
The development of an inflatable apple-picking robotic arm addresses labor shortages in the fruit-picking industry, offering a potential solution to reduce waste from unharvested crops.
- A Washington State University-led team has developed an inflatable robotic arm designed to pick apples without harming trees or branches.
- The robotic arm weighs less than 50 pounds and uses sensors to identify and collect apples, completing the process in about 25 seconds.
- Currently, the robotic arm is slower than human pickers, who can harvest an apple every three seconds, but improvements are being explored to enhance its efficiency.
- The technology is aimed at addressing labor shortages in Washington state, which produces over $2 billion in apples annually and faces challenges with an aging workforce.
- The team plans to patent the robotic arm technology and potentially lower the price below its current $5,500 cost.
The apple-picking robotic arm is part of evolving robotics aimed at improving agricultural efficiency amidst labor shortages. Washington state is the leading apple-producing state in the U.S. Future developments will focus on more tasks for the robot, including pruning and spraying.
Read the original ↗Artificial Intelligence (392)
The development of AI agents in code sandboxes can enhance testing and deployment, which is critical for advancing artificial intelligence applications.
- The article identifies five code sandboxes designed for AI agents.
- These sandboxes facilitate the development and testing of artificial intelligence technologies.
Code sandboxes provide isolated environments for developers to experiment without affecting production systems. They are essential for testing AI functionality in controlled settings. The growth of AI technologies continues to drive innovation in various fields.
Based on the headline only — the full article could not be fetched.
Read the original ↗
ArenaRL: Scaling RL for Open-Ended Agents via Tournament-based Relative Ranking
The development of ArenaRL could significantly enhance the performance of AI agents in open-ended tasks, enabling more effective solutions in complex, real-world scenarios.
- Reinforcement learning for large language model (LLM) agents faces challenges in open-ended tasks due to discrimination collapse from scalar scoring.
- ArenaRL proposes relative ranking and pairwise evaluation mechanisms to improve the reliability of performance assessments.
- Empirical results indicate that ArenaRL achieves high accuracy in advantage estimation while maintaining efficient computational complexity.
- New benchmarks, Open-Travel and Open-DeepResearch, were created to support the evaluation of LLM agents in diverse settings.
ArenaRL represents a new paradigm in reinforcement learning focused on enhancing LLM agents' abilities in tasks with numerous potential solutions. Previous reinforcement learning approaches often failed to provide effective evaluations due to lack of objective standards. Upcoming developments will likely focus on refining this methodology and expanding its applications in real-world environments.
Read the original ↗
Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
The introduction of conditional memory in Transformer models has the potential to significantly enhance reasoning capabilities and efficiency in large language models, impacting AI applications across various fields.
- Conditional memory, enabled by the Engram module, enhances Transformer models by optimizing knowledge lookup and improving reasoning abilities.
- Engram scales to 27 billion parameters, outperforming a comparable Mixture-of-Experts baseline in tasks requiring knowledge retrieval.
- The implementation of conditional memory shows performance improvements in reasoning-centric benchmarks, including gains in MMLU and ARC-Challenge.
- Engram's architecture allows the separation of memory and reasoning, which aids local dependency management and enhances long-context retrieval efficiency.
Transformers are a type of AI model used for various language tasks, struggling with efficient knowledge retrieval without dedicated lookup mechanisms. Previous methods relied on computational simulations, while this new approach presents a scalable solution. This development heralds a shift towards optimized memory use in AI, which may inform future model designs.
Read the original ↗