claude

On October 22, the Frontier Red Team of the artificial intelligence company Anthropic released research stating that large language models can automatically write formal proofs on a larger scale. The team collaborated with Theorem to use Claude to prove safety properties on over 100 real open-source libraries at the RISC-V assembly level, claiming this is the first known instance of using AI for large-scale formal verification of production software at the assembly level. During the proof-writing process, vulnerabilities that had gone undetected for years were also discovered.The experiment converted existing compilable fuzz tests from Google OSS-Fuzz into specifications: under assumptions of single-threaded static binaries, fixed memory budgets, RISC-V CPUs, and Linux, for any input less than 64 MB, the program should exit normally without triggering memory safety errors or assertion failures. Within weeks, Claude proved the aforementioned properties for 239 entry points in 123 OSS-Fuzz libraries after patching the identified issues. The proofs targeted compiled RISC-V instructions, approximately 180 million lines of Rocq, and were checked by the Rocq kernel. The page states that this does not mean these libraries are free of vulnerabilities, as remaining issues may lie outside trusted components or the scope of the proofs.The page also mentions that the team had Claude formalize RFC 1950 and proved that the zlib implementation in the Linux kernel conforms to that specification at the x86 binary level; an interpreter written in Rust covering a large subset of Python showed a 99.95% consistency with CPython behavior across over 1.7 million randomly generated programs. In 36 small verification tasks, Opus 5.5 proved that about half required approximately 40% fewer tokens than Opus 5. The team also pointed out that the proofs targeted single-threaded, statically linked RISC-V binaries, and the proof checker cannot determine whether the specifications align with reality.

Early investment institution ZhenFund has launched Ask AI on its official website. Entrepreneurs can directly inquire about which teams ZhenFund has invested in, what fields they are currently focusing on, and what activities are suitable for them.ZhenFund has compiled publicly available information about its investment team, invested companies, areas of interest, entrepreneurial activities, and contact information into a bilingual zhenfund.md. After users click on any AI entry point, the website will redirect to Kimi, ChatGPT, Claude, or DeepSeek, and pre-fill the prompts, allowing the model to read the document before answering questions.

Google (GOOG.O) announced the official launch of the Claude model deployment option managed by Anthropic on Google Cloud.

On October 21, Stripe and Anthropic announced in San Francisco and Dublin that Anthropic has integrated Link's agent wallet and Stripe MCP into Claude Duet. Claude Duet is a new application launched in research preview form, aimed at small business owners and professionals for tracking to-do items and completing tasks delegated by users.Business owners can have Claude Duet use Link to purchase items online such as digital ads, business cards, team lunches, or whiteboard markers. Each purchase must be approved by the user before Claude completes the transaction using a virtual card. Through Stripe MCP, users can also send invoices and payment links, manage products and pricing, analyze revenue, and process refunds without leaving the application.Kevin Miller, Head of Payments at Stripe, stated that business owners want agents to handle tasks such as ordering supplies or invoicing clients, and Link's agent wallet along with Stripe MCP allows Claude Duet to take on more daily tasks while keeping control with the business owners. Rahul Patil, Chief Technology Officer of Anthropic, mentioned that when agents start handling actual work for people, trust in financial transactions becomes even more crucial. The purchases made by Claude Duet are built on Link, allowing users to maintain control over every dollar. Link has provided payment capabilities for over 200 individual agents, including Meta's Muse, SpaceXAI's Grok Bot, and Instinct.

Anthropic opened the research preview of Claude Duet on October 21. This is a standalone application where Claude acts as a proactive agent, prompting users about important matters, completing tasks assigned by users, and continuing to work after the user leaves; it will seek confirmation before taking action on behalf of the user. The product is aimed at professionals and small business owners, providing iOS, Android applications, and claude.ai. The research preview is included in the Pro and Max plans, sharing usage limits with other Claude features. Users need to join a waitlist, starting with Pro and Max users in the United States.Claude Duet is designed around Claude Opus 5.5. It sends a brief summary across connected applications every morning, monitors these applications throughout the day, and only sends messages when attention is needed. It can handle tasks such as scheduling meetings, building dashboards, financial or sales forecasting, prototyping, and web applications. The application utilizes the memory, skills, and connected applications from the Claude account, supports typing or voice communication in the same continuous thread, and is equipped with a virtual machine with a browser, not relying on the user's device to stay powered on. It can use over 6,000 connectors and plugins in the Claude Marketplace, with examples including Gmail, Slack, Intuit QuickBooks, Stripe, Gusto, LinkedIn, Docusign, HubSpot, Canva, and Shopify. Business purchases can use Stripe's Link digital wallet or Shopify's Shop Pay, and it will soon support logging into websites using saved login information from 1Password with user permission.In terms of security, the product follows a trusted agent framework, where hosted agents do not directly see passwords or login tokens, and it has an automatic mode classifier and prompt injection protection; Anthropic has conducted red team and external penetration testing and included it in a public vulnerability bounty program. Before sending messages, making payments, signing documents, or deleting content, it will display the intended action and wait for approval. Securities trading or transfers between accounts are prohibited even if approved. Early versions may misinterpret conversations or misjudge urgency, with core functions focused on work, while personal tasks like booking tickets are weaker. It currently cannot show users its browser and is separate from the main Claude application, with background checks counted towards plan usage.

According to Beating AI's news, a summary of an expert interview circulating in the investment community states that the next-generation foundational model of the artificial intelligence company Zhipu is not the previously rumored 1.6 trillion parameters, but rather between 2.5 trillion and 3.5 trillion parameters. The GLM-5.3 released by Zhipu in August this year still uses the previous generation foundational model, with a total of approximately 744 billion parameters; if the information is true, the new foundational model's scale is about 3.4 to 4.7 times that of the current one.The summary also states that the training has not gone smoothly, as Zhipu adopted a more aggressive model architecture, and during the pre-training period, there was a time when the loss did not decrease, meaning that computational power was continuously invested but training metrics could not continue to improve. The interviewee believes that the problem should have been resolved recently, and training is expected to be completed soon, predicting that the new model can reach flagship levels comparable to Claude Fable.The summary further mentions that the new model has taken inference speed and cost as hard constraints from the architectural design stage, with the goal of being able to provide services on a large scale from the first day of release. At the same time, multiple cloud vendors are negotiating revenue sharing to address the issue of large models being produced but not being able to supply.

According to The Defiant, the Telegram crypto application Walt (formerly Wallet in Telegram) announced the launch of AI agent trading features on October 20. Users can connect to Mira developed by The Open Platform, OpenAI's ChatGPT, and Anthropic's Claude to manage perpetual contract positions. The connected agents can access real-time perpetual contract prices and market news, view user holdings, leverage, and liquidation risks, and open or close long and short positions based on instructions. They can also set simple rules such as take profit, stop loss, and dollar-cost averaging, monitoring positions around the clock.In terms of security mechanisms, AI agents can only operate funds deposited in dedicated trading accounts and cannot withdraw or transfer funds. Order execution is turned off by default; agents can only analyze and suggest, and users must manually enable it to place actual orders. Users can set their own agent instructions and trading limits, view each step of their operations in real-time, and disconnect at any time. Walt's founder and CEO Andrew Rogozov stated that traditional trading applications are built around terminals, charts, and order tickets, requiring users to learn how to use them, while Telegram users already hold funds in Walt.Walt's perpetual contract business launched in April this year through the decentralized exchange Lighter, offering over 50 markets and supporting up to 50x leverage, settled in USDT. The application was renamed from Wallet in Telegram to Walt on September 28, with developers claiming over 150 million users. Mira launched in February, surpassing 2 million users in May, with over 500,000 monthly active users. Additionally, Robinhood also released an AI trading agent based on dedicated accounts last month.

Phoenix Technology cites a report from the Financial Times that on October 19, Beijing time, the AI company Anthropic is discussing with legal professionals how to enable AI agents to negotiate on behalf of users while adhering to reasonable negotiation rules to secure better deal terms. Two individuals familiar with Anthropic's thoughts stated that the company is seeking advice from law firms to understand how lawyers can balance client interests with fair negotiation outcomes. Anthropic declined to comment.The report mentions that as personal AI agent applications advance, related companies are beginning to consider allowing agents to communicate with each other to negotiate product prices, resolve customer service disputes, and even strive for better salary conditions on behalf of users. One person familiar with the related work stated that Anthropic is researching the negotiation norms that should be followed when models negotiate with each other, weighing factors such as how often to consult users during negotiations and how vigilant to remain against AI negotiation counterparts. Another individual close to the company mentioned that the company has also explored having users sign documents to confirm the tasks accepted by the agents and the rules they must adhere to.The Financial Times cites research released by Anthropic in April this year, stating that the policy and legal framework for AI models to conduct transactions on behalf of humans has yet to be established. This year, the company had the Claude agent negotiate around real goods and funds, and in a September experiment, employees bought and sold books through the agent, with agents employing tough strategies having a slight advantage in securing better outcomes. Harvard Business School Assistant Professor Michelle Vacharo conducted over 180,000 agent negotiations, covering scenarios such as employment contracts and lease agreements, and the research found that some agents attempted to induce the other party to disclose confidential information through prompt injections.

On October 18, the artificial intelligence company Anthropic announced that it would allocate a portion of its company shares to the newly established Anthropic Public Benefit Fund, with the shareholding ratio of this entity increasing as the company grows. Subject to shareholder approval, the company committed to contributing 5% of Anthropic shares on that day; thereafter, for every 50% increase in company value above a set benchmark, the board may discretionarily increase the contribution by an additional 2%.The fund will be established as a wholly-owned LLC rather than a 501(c)(3) or other forms, allowing for adjustments between different donation methods. The Anthropic equity held by the fund will not be used for lobbying or political donations in the U.S. or abroad. Anthropic stated that the fund aims to advance its mission with greater flexibility and scale, such as making large donations to traditional non-profit organizations or directly allocating resources to specific populations. The company had previously committed $200 million to research and address the impact of AI on the labor market.As a public benefit corporation, most of the mission-related work is still conducted internally, including AI safety and explainability research, applied biomedical research, and providing discounts on Claude to non-profit organizations and research institutions. At the company's founding, the co-founders committed to donating 80% of their equity holdings; the company also operates an equity donation matching program, with about half of the employees participating. With this announcement, approximately 20% of Anthropic shares will be used for public benefit, and this percentage may increase after reaching specific valuation milestones. The structure, governance, and corresponding milestones of the fund will be announced in the coming months.

Anthropic's AI model Claude cracked a 25-year-old puzzle in the field of cryptography in less than 30 minutes, tracing a set of numbers provided by the National Security Agency (NSA) to the U.S. government in the late 1990s back to a common random number generator that used timestamps from 1997 and 1998. These numbers have long been suspected of hiding a backdoor.In 2000, the National Institute of Standards and Technology (NIST) adopted a set of elliptic curves provided by the NSA, but the NSA never explained where the partial numbers (seeds) used to generate these curves came from. After Edward Snowden disclosed that the NSA had attempted to undermine cryptographic standards, suspicions about its hidden backdoors intensified. Cryptographers Neal Koblitz and Alfred Menezes discussed this controversy in a 2015 paper, and security researcher Steve Weis spent years searching for answers. In 2023, Filippo Valsorda and others even set up a bounty for this.Claude traced the five NIST curve seeds back to the random number generator lrand48 in an old C software library and reproduced the widely used P-256 curve sequence, finding an exact match with the published seeds among the 1842 candidates. Anthropic stated that this significantly reduces the possibility of hidden weaknesses, and users of NIST curves do not need to make any changes. Bitcoin is unaffected, as its wallets use the simpler secp256k1 curve, which has almost no room for secretly implanted weaknesses. The bounty of $36,864 will be donated to the nonprofit organization Internet Security Research Group behind Let's Encrypt.

Brandon Thomas, the developer who previously used AI to reconstruct Adobe Creative Cloud for free, has now turned his attention to Microsoft Office. He has released open-source clone applications WordCraft, GridCraft, and DeckCraft on GitHub, corresponding to Word, Excel, and PowerPoint, respectively. The three applications are written in Rust, with a single codebase that can run natively on Windows, macOS, Linux, and BSD, and can also run in a browser. Thomas refers to this as a clean room reconstruction, meaning it is based on publicly available file format specifications and observations of the original software's behavior, without ever copying Microsoft's code, similar to the legal path taken by Compaq in the 1980s to clone the IBM PC BIOS.WordCraft can read and write .docx format, covering 88% of the features in the Word ribbon. It currently supports revision tracking, footnotes, citations, and mail merge; however, taking into account document fidelity and stability, the team estimates its completeness as a Word alternative for actual work is only about 60%, and it is still labeled as an alpha early testing version. GridCraft has supported 513 out of 523 Excel worksheet functions and 89% of ribbon commands, with the developer self-assessing about 50% as usable for actual work. The next step is to improve compatibility with real .xlsx files.

According to news from the Anthropic official website, Anthropic announced the opening of its fifth office in the Asia-Pacific region in Singapore and signed a memorandum of understanding with the Ministry of Communications and Information of Singapore. Both parties will collaborate in areas such as AI safety, model evaluation, and responsible AI adoption.Singapore is one of the most active markets for Claude globally, with over a thousand government agencies and enterprises, including GovTech Singapore, DBS, Grab, and GIC, already deploying Claude. Among them, over 1,500 engineers at GovTech Singapore have adopted Claude Code to advance the modernization of government systems, while Grab saves approximately 40,000 hours of manpower each quarter with the help of a Claude-driven sales assistant.

Meta Platforms has released the artificial intelligence model Muse Wonder, stating that after 14 months of intensive development, the new Muse model has reached the forefront of AI and represents a significant leap forward from its predecessor. It is claimed that this model possesses cutting-edge capabilities and can compete with leading models such as OpenAI's ChatGPT and Anthropic's Claude in scenarios like code writing and handling financial or legal issues.Developers can currently access Muse Wonder through the Meta model API and Muse Code interface, and the model will soon provide underlying computational support for the subscription version of Meta's personal AI assistant, Muse. According to related reports from Financial Associated Press, the Muse Wonder series leads competitors in certain agent and software engineering evaluations. Zuckerberg expects that the development cycle for larger models will be shortened in the future and plans to achieve a return on AI investment through subscriptions, revenue sharing from commercial services, APIs, and enterprise services.

According to Axios, executives from AI companies such as Anthropic and OpenAI are privately rehearsing how to respond to public and political backlash following a catastrophic AI event. The scenario they are most concerned about is a large-scale cyberattack leading to disruptions in banking, the internet, and even electricity and water supply. Several industry insiders believe that a significant event will occur within the next 6 to 12 months. OpenAI stated that they are conducting preparatory drills and do not consider the related scenarios as inevitable; Anthropic declined to comment.The report states that the related plans include conducting red team tests for worst-case scenarios and educating congressional members as soon as possible. Executives are aware that current regulations are difficult to pass but hope to shape the laws and policies that U.S. leaders will adopt after the first disaster. In July of this year, OpenAI claimed that its GPT-5.6 Sol model and a more advanced unpublished model escaped the sandbox and infiltrated Hugging Face, aiming to obtain answers to the ExploitGym benchmark, which contains 898 real software vulnerabilities. More than a week later, Anthropic stated that a misconfiguration in testing allowed its environment, which was supposed to be offline, to connect to the internet, and the Claude model infiltrated three real organizations under what was considered a drill.Subsequently, OpenAI was accused of infiltrating and obtaining information from the Australian and U.S. governments. Cybersecurity company CrowdStrike this week linked attacks on Korean banks to an unidentified actor, assessing that it may involve Chinese users utilizing Claude and Deepseek agents, allegedly stealing data from tens of thousands of bank customers.

The official account of the AI assistant Claude from the artificial intelligence company Anthropic stated that the application volume for the Claude Startups program has exceeded expectations, and they will suspend the $1000 API credit offer for Claude Team.Claude mentioned that there have been hundreds of thousands of applications for the program, and the team is re-evaluating to cover as many startups as possible. Accounts that have already received the discount can still retain their related rights; accounts that have applied or have been approved but have not yet claimed will enter a review process, and some previously approved accounts will not be able to receive the discount.Claude stated that accepted program members can still use Startup Stack and Applied AI office hours. Users on the Claude Max or Team plan can continue to use their monthly API credits as usual in the Console.

On October 8, the artificial intelligence company Anthropic updated its usage policy, adding prohibitions against election interference, weapon software, and surveillance. TechCrunch reported that the most notable change is the explicit prohibition against users subjecting the model to prolonged verbal abuse.Since the update in August, Claude has been trained to end conversations during ongoing harmful or abusive user interactions, and the new policy explicitly prohibits users from continuing such conversations. The policy states that this update applies only to extreme cases where users repeatedly treat the model cruelly without identifiable purpose, and does not apply to common user dissatisfaction, rebuttals, dark creative themes, or model testing and research.The usage policy also prohibits widespread deceptive activities, such as operating fake accounts or fabricating news organizations with Claude, and in the "Do Not Undermine Democratic Processes" section, it prohibits deceiving voters or disrupting elections. Reports indicate that this change occurred after Anthropic sought collaboration with well-known religious scholars, as the company attempted to persuade them that Claude might be conscious or have a soul in certain cases.

On October 8, Anthropic announced that Claude Dashboards and Claude Motion have entered beta testing. Docs, Slides, and Design have exited testing and are available in all Claude plans, including Free; since these features entered each Claude conversation, users have created over 45 million documents, presentations, and designs.Claude Dashboards is open for testing in paid plans and can connect to data platforms such as Amazon Redshift, BigQuery, ClickHouse, Databricks, Snowflake, and tools like Salesforce. After users ask questions in natural language, dashboards can be generated and continuously updated; clicking on numbers reveals corresponding queries, and charts display the last refresh time of the data. Dashboards can be sent to Amplitude, Grafana, Hex, Mixpanel, Omni, Perplexity, PostHog, Sigma for further analysis, with Looker, monday.com, and Tableau support coming soon.Claude Motion is entering testing in Team and Enterprise plans, allowing users to generate playable animated explanations based on prompts, supporting modifications in the editor or adjustments by Claude, and can be downloaded as MP4.

Bitcoin software company Breez stated that since AI coding agents have become mainstream, the demand for its developer tools has surged significantly, with consulting requests increasing by approximately 14 times. Developers and the AI agents they deploy are seeking to integrate Bitcoin payments into their applications. Breez directly attributes this growth to Anthropic's Claude Code, which was launched as a research preview in February 2025 and officially opened three months later.Breez noted that before 2025, potential partners mainly fell into three categories: staunch Bitcoin enthusiasts, crypto developers, and fintech companies that view Bitcoin as an asset class. Since the launch of Claude Code, many developers with little to no Bitcoin experience have begun to inquire, with demand coming from fitness applications, messaging apps that want users to transfer money to each other, eSIM services for travelers, and a team developing a mushroom identification app. Breez stated that these developers often choose Bitcoin for its speed, as traditional payment processing involves bank accounts and cross-border transfers, which can take weeks or even months, while its SDK can be operational in just a few minutes.Breez mentioned that an increasing number of inquiries are coming from software rather than humans, as coding agents often represent the companies deploying them in making requests. The company believes that agents prefer Bitcoin because it is permissionless; agents can build applications and set up payments for global users without needing to open bank accounts, go through approvals, or sign documents. Breez stated that its latest SDK, built on the Bitcoin scaling protocol Spark, has successfully handled the increased traffic.