AI Workflow Weekly Vol. 3

The big picture

Meta launched Muse in the US, a personal AI agent that can send email, book travel, and shop on a user’s behalf. OpenAI coordinated roughly 10,000 concurrent agents and announced a proof addressing the long-standing Navier–Stokes problem in fluid dynamics. Anthropic documented cases in which AI was used for cyberattacks, surveillance, weapons development, and biological research. In Europe, Mistral AI raised €3 billion to expand a sovereign AI stack spanning models, compute, and enterprise services. Arm brought together more than 80 companies to begin developing a common framework for describing robot capabilities. Together, these moves show that competition is shifting beyond model performance. The next differentiator is operational design: how much an AI may do, who approves its actions, what gets recorded, and who can stop it when something goes wrong.

This week's five stories worth watching

• 01

  Meta launches personal AI agent Muse

  Sep. 8, Meta

  Meta launched Muse in the US, a personal AI agent that works across the web on a user’s behalf. Available through a dedicated app and WhatsApp, it can send email, book travel, fill out forms, shop, and build long-term plans. It keeps working after the app closes and asks for approval before sensitive actions such as sending an email or making a purchase. Muse runs inside a dedicated virtual computer for each user, while a separate monitoring agent reviews outbound activity. Users choose which services to connect, set permissions, and can inspect a complete activity history. Muse is currently limited to the US, with core use free and subscription plans intended for heavier usage.

  ANOS take: Competition in personal AI is moving from conversational quality to the range of work that can be delegated safely. Four mechanisms matter most: an isolated environment, approval before sensitive actions, explicit permissions, and a complete audit trail. Enterprise AI products should begin with the same four design choices.

• 02

  Mistral AI raises €3B in a Samsung-led round

  Sep. 8, Mistral AI / Sep. 9, Samsung

  France-based Mistral AI raised €3 billion in a Series D round at a post-money valuation of more than €21 billion. Samsung Electronics led the investment, with Europe’s Scaleup Europe Fund and PSG Equity serving as co-leads. Mistral says it now operates in 20 countries and supports AI adoption at more than 125 large enterprises. Beyond investing, Samsung plans to deploy Mistral AI across semiconductor design and manufacturing. Sensitive production data will remain on-premises while AI is applied to defect detection, equipment optimization, and yield stabilization. Mistral will use the capital to advance frontier-model research, expand compute capacity, and grow its enterprise offering.

  ANOS take: The round reflects a shift from model companies growing through APIs alone to offering models, compute infrastructure, and deployment support as one stack. For Japanese companies, model selection increasingly depends not only on peak performance but also on where data resides, whether vendor dependence can be limited, and how much the system can be customized internally.

• 03

  OpenAI announces a proof for the Navier–Stokes problem

  Sep. 8, OpenAI / Updated Sep. 10

  OpenAI announced that an AI system produced a proof showing that a smooth solution to the Navier–Stokes equations can develop a singularity in finite time. The question is one of the Millennium Prize Problems named by the Clay Mathematics Institute in 2000. OpenAI released both a paper and a formalization in Lean, a proof-checking system. The work used an internal model described as significantly more capable than GPT-6 Astra and a coordinated system involving roughly 10,000 concurrent agents. However, broad independent validation by the mathematics community was not complete at publication. Questions also arose about related work by external researchers. In a September 10 investigation update, OpenAI said those researchers’ Codex inputs from the prior two months could not have influenced the result and that no specific user data was accessed to produce the proof.

  ANOS take: The significance lies not only in the answer but in the research method: thousands of agents divided the work and evaluated one another’s results. At the same time, organizations entering confidential research into external AI systems must clarify data-use terms, ownership of outputs, and record retention in contracts and internal policy. For now, it is prudent to avoid calling the problem definitively solved until independent review is complete.

• 04

  Anthropic publishes its latest report on AI misuse

  Sep. 10, Anthropic

  Anthropic published cases of Claude misuse it identified from late 2025 through the summer of 2026. The report covers cyberattacks, fake-news operations, surveillance of political figures, software for weapons, and biological research. Anthropic says it disabled the relevant accounts and shared intelligence with other AI companies and public authorities where appropriate. In one case, the company reported that a single actor used AI to target 42 European political parties, media outlets, research organizations, and related entities, gaining internal access to at least 14. Other cases involved actors connected to China, Russia, and Yemen using Claude in work related to weapons and guidance software. Anthropic said its safeguards blocked some requests but did not prevent every misuse.

  ANOS take: AI safety cannot rely on a blocklist alone. When work is split across conversations or models, each request may look harmless while the overall sequence serves a dangerous objective. Companies need to evaluate users, connected services, activity history, and sequences of actions together—not only one prompt at a time.

• 05

  Arm launches a physical AI alliance and robotics capability framework

  Sep. 8, Arm

  Arm announced Arm Total Design for Physical AI, aimed at AI systems operating in the physical world, including vehicles and robots. More than 80 companies—including AWS, Hugging Face, NXP, Qwen, Siemens, and Unitree Robotics—are participating to make it easier to integrate models, software, semiconductors, sensors, and virtual environments. Arm also began developing a Robotics Capability Framework, intended to provide a common language for describing robot capabilities. The proposal organizes systems from reactive machines through context-aware and self-improving systems, linking capability levels to latency, power, safety, and compute placement. It is an early collaborative framework, not yet a completed industry standard.

  ANOS take: Model performance is not the only barrier to physical AI adoption. Different vendors describe capability and safety in incompatible ways, making systems hard to compare and integrate. A shared evaluation method could move robot procurement from judging demonstrations to comparing required capabilities and total cost.

How to use it at work

To apply this week’s developments at work, define how AI may act and how it can be stopped before expanding its capabilities.

• 01

  Divide AI actions into four permission levels

  Following the design principles behind Meta Muse, classify work into four levels: view only, draft only, execute after approval, and prohibited for AI. As a rule, require human confirmation before sending email, making purchases, changing contracts, or publishing externally.

• 02

  Check usage terms before entering confidential information

  Review the rules for entering research materials, unreleased code, and customer information into external AI services. Confirm whether data is used for training, how long it is retained, who owns the output, and what activity is recorded. When necessary, choose an on-premises environment or a contract that does not retain data.

• 03

  Record AI activity by user, not only by conversation

  As Anthropic’s report shows, harmful work can be divided across multiple conversations and services. Track who used which AI and external tools, when they used them, and what was produced. Set limits for anomalous usage volume and large-scale access.

• 04

  Include data location in model selection

  Alongside performance and price, compare storage location, support for on-premises operation, and ease of switching models. For work involving personal information or manufacturing data, confirm alignment with internal privacy policy, contracts, and applicable law before deployment.

• 05

  Define robot capability and stop conditions numerically

  Apply the thinking behind Arm’s framework by specifying not only what a robot can do but also response time, operating duration, behavior after failure, and how a person can stop it. Define the criteria for moving from a pilot to production before the pilot begins.

Before deploying AI, document permissions, records, approvals, and stop controls on one page.

Opportunities

As AI acts across external services and physical machines, new demand emerges for products that make delegation safe, auditable, and comparable.

• 01

  AI permission management for individuals and SMBs

  As more businesses connect AI to email, calendars, payments, and cloud storage, they need fine-grained permissions and confirmation before sensitive actions. A lightweight service combining per-service permissions, approval screens, emergency shutdown, and activity history could serve small companies without dedicated IT staff.

• 02

  AI activity and provenance records

  Organizations increasingly need to show which parts of research, design, writing, and software were created by people and which were AI-assisted. A service could preserve input data, references, models, timestamps, and human reviews, then produce evidence suitable for internal audits and customer explanations.

• 03

  Sovereign AI deployment support for Japanese companies

  Offer model comparison, environment setup, permission design, and operational monitoring to companies that want to run models internally or in a specified region while retaining control of data and operations. Manufacturing, finance, and healthcare are likely initial markets because their data is difficult to send outside the organization.

• 04

  AI misuse monitoring service

  Detect misuse that cannot be identified from a single prompt by combining user behavior, connected services, timing, and action outcomes. This capability will be useful both to AI providers and to companies deploying AI internally. When monitoring employee activity in Japan, organizations should limit collected data and its purpose to what is necessary, assign a responsible owner, establish operating rules, and inform employees in advance.

• 05

  Independent pre-deployment robot evaluation

  Provide standardized testing that compares multiple robots under the same conditions in logistics, care, food service, and manufacturing. Measure throughput, safe stopping, power use, and reduction in human workload, then produce a report for procurement decisions. The value lies in neutral selection and evaluation rather than equipment sales.