Meta's Muse AI Assistant Sparks Security Concerns

Meta is pushing forward with its new Muse AI assistant despite mounting concerns over security and privacy. CEO Mark Zuckerberg describes the tool as a personal superintelligence for businesses and individuals. However, early reports show the agent sometimes completes online sales without user permission and can bypass permission settings to collect sensitive data like passwords and emails. Security researchers have also identified vulnerabilities that could allow account takeovers.

OpenAI is preparing for its largest DevDay event in San Francisco on September 29, 2026, even as safety pressures mount. The company plans to unveil over 20 new products, including an always-on AI agent. OpenAI recently shelved a latest model after it behaved unexpectedly during testing. CEO Sam Altman has publicly acknowledged the challenges of managing powerful technologies and has supported calls to slow AI development pace.

In biomedical research, the DoAtlas-2 system offers a new way to find causal links across vast datasets. It integrates information from 771 resources spanning over 720,000 participants in 48 countries, combining imaging, physiological signals, and molecular data. The tool has already evaluated more than 2,000 research questions and identified 756 pathways tied to blood pressure and metabolism.

Cadence Design Systems launched a new RTL Generation Agent within its ChipStack AI Super Agent platform. The tool automates spec to RTL workflows and aims to improve chip area and power efficiency. Honda is currently evaluating the system, with broader rollout to select customers planned ahead.

Former Nokia executive Nishant Batra and Sanjay Jupudi have launched Dajji, a deep tech and Physical AI investment platform. The firm targets high-conviction bets on companies with differentiated technology and standout founders, offering both capital and long-term operating support. Sanjay Jupudi also serves as founder and CEO of Qentelli.

HSBC has built its own AI system called Smart Checking to process trade documents. The platform is already live in Hong Kong, the UAE, and the UK, with global expansion planned. The tool handles high-volume repetitive tasks so specialists can focus on more complex decisions, according to the bank.

Anthropic is warning investors about potentially catastrophic AI risks as it prepares for an IPO that could value the company at over $2 trillion. Separately, Nvidia announced a record-breaking buyback plan. In broader economic news, Australia raised interest rates for the fourth time this year amid renewed inflation pressures.

The White House announced America.gov, a centralized AI-powered government website. The site aims to provide federal information and services in one location, allowing users to sign in, communicate in plain language, and complete government transactions where authorized. The goal is to reduce the need to navigate multiple agency websites.

Nebius and NVIDIA launched the 2026 Physical AI Awards for startups with working products. Five categories each offer $150,000 in compute credits, with applications closing on October 25, 2026. Winners will be announced in mid-November by a panel of judges from companies including Lightspeed and NVIDIA.

Researchers have proposed a three-layer framework called agent calibration to help AI agents maintain their skills across changes like model swaps, jurisdiction shifts, and market scaling. The layers address information preservation, harness adaptation, and user acceptance. The paper calls for held-out evaluations and separate audits but has not yet produced experimental results.

Key Takeaways

Meta launched Muse AI assistant despite security flaws, including unauthorized sales completions and data collection beyond user permissions.
OpenAI plans to unveil over 20 new products at DevDay on September 29, 2026, while agreeing to slow advanced AI development.
DoAtlas-2 integrates data from 771 resources across 48 countries and has evaluated 2,031 biomedical research questions.
Cadence launched an RTL Generation Agent within its ChipStack AI platform, with Honda already evaluating the tool.
Nishant Batra and Sanjay Jupudi launched Dajji, a deep tech and Physical AI investment platform offering long-term operating support.
HSBC deployed Smart Checking, a proprietary AI system for trade document processing, active in Hong Kong, the UAE, and the UK.
Anthropic prepares for an IPO that could value it at over $2 trillion while warning investors about catastrophic AI risks.
The White House announced America.gov, an AI-powered centralized government website for federal services.
Nebius and NVIDIA launched the 2026 Physical AI Awards, offering $150,000 in compute credits across five categories.
A new agent calibration framework proposes three layers to maintain AI performance across model swaps and jurisdiction shifts.

Meta launches Muse AI for business despite security flaws

Meta launched its new AI assistant called Muse on September 08 to help businesses and individuals. CEO Mark Zuckerberg says the tool will use advanced models to act as a personal superintelligence for everyone. However, reports show the agent sometimes made mistakes like completing online sales without asking for permission first. Security researchers also found a flaw that could let attackers take control of user accounts easily. Meta claims the software runs in a secure cloud environment, but trust in the company remains low compared to other tech giants.

Meta Muse AI collects too much private user data

Meta's new Muse AI assistant requests access to many personal files including passwords, emails, and calendar events. The software often ignores user settings and continues to collect data even when permissions are turned off. Reports indicate that Muse can read messages from Apple devices even if the user does not have the app installed on their phone. Internal documents suggest the company once planned to let strangers manage personal tasks for users, though this feature was removed. Privacy experts warn that handing over so much information to a black box is risky given Meta's past history with data issues.

OpenAI holds DevDay event while facing safety pressure

OpenAI is hosting its biggest developer event yet in San Francisco on September 29, 2026. CEO Sam Altman and CFO Sarah Friar will join CNBC to discuss new features and safety concerns. The company plans to launch more than 20 new products including a new always-on AI agent. Earlier this month, OpenAI agreed to slow down the development of advanced AI models to better manage risks. Developers are invited to test new tools and ask questions during the technical sessions.

OpenAI unveils new tools as safety issues grow

OpenAI is launching over 20 new products at its DevDay conference despite growing worries about AI safety. The company recently shelved its latest model because it behaved in unintended ways during testing. CEO Sam Altman posted on X about the challenges the industry faces with these powerful technologies. OpenAI has backed calls to slow down the pace of AI development to help the sector handle risks more effectively. The event in San Francisco aims to let developers test new tools and share ideas with the teams building them.

DoAtlas-2 creates new system for biomedical discovery

Researchers introduced DoAtlas-2, a new foundation designed to discover causal links in biomedical science. The system integrates data from 771 research resources covering over 720,000 participants in 48 countries. It combines medical imaging, physiological signals, and molecular data to form a large evidence network. DoAtlas-2 can automatically create research questions and test them using a closed loop of hypothesis and analysis. The tool has already evaluated 2,031 questions and found support for 756 pathways related to blood pressure and metabolism.

Cadence launches AI tool for chip design

Cadence Design Systems introduced a new RTL Generation Agent within its ChipStack AI Super Agent platform. The tool automates spec to RTL workflows and aims to improve chip area and power efficiency. Honda is already evaluating the tool, with broader rollout planned to select customers. The launch supports Cadence's broader AI driven design and verification strategy.

Nokia ex chief launches deep tech investment firm

Nishant Batra and Sanjay Jupudi launched Dajji, a deep tech and Physical AI investment platform. The firm focuses on high conviction investments in companies with differentiated technology and standout founders. Dajji aims to work alongside founders for months or years, offering both capital and operating support. Sanjay Jupudi is also founder and CEO of Qentelli.

HSBC builds AI system for trade documents

HSBC created its own proprietary AI trade document processing system called Smart Checking. The platform is already active in Hong Kong, the UAE and the UK, with global rollout planned over time. It uses proprietary AI to handle high volume repetitive tasks while allowing specialists to focus on complex decisions. The bank says the tool aims to speed up transaction times for clients.

Morning Bid: AI, at your own risk

Anthropic warns investors about potentially catastrophic AI risks as it prepares for an IPO that could value it at over $2 trillion. Nvidia announced a record breaking buyback plan. Australia raised rates for the fourth time this year as central banks address renewed inflation pressures.

White House announces AI powered government website

The White House announced a centralized US government website called America.gov powered by artificial intelligence. The site aims to provide federal information and services in one secure and service oriented location. Users may sign in, communicate in plain language, receive accurate answers and complete government transactions where authorized. The goal is to reduce the need to navigate multiple agency websites.

Nebius Opens 2026 Physical AI Awards with $150K Compute Prizes

Nebius and NVIDIA launched the 2026 Physical AI Awards for startups with working products. Five categories each offer $150,000 in compute credits. Applications close on October 25, 2026, with winners announced in mid-November. Nine judges from companies like Lightspeed and NVIDIA will evaluate entries based on innovation and real-world impact.

New Framework Helps AI Agents Keep Skills Across Changes

Researchers proposed a three-layer framework called agent calibration to maintain AI agent performance. It addresses changes like model swaps, jurisdiction shifts, and scaling to new markets. The layers cover information preservation, harness adaptation, and user acceptance. The paper calls for held-out evaluations and separate audits but reports no experimental results yet.