RWA platform Midas launches tokenized investment products mWIN and mGLOBAL on Avalanche.

Dongcha Beating AI News Flash, November 5 — A developer complained on GitHub that Claude Code leaked his email, leading to dozens of scam emails hitting his inbox daily. Anthropic engineer Lydia Hallie explained that earlier versions of Claude often randomly inserted author names when generating code. To resolve this, the team added users’ login emails to the model’s context window. When Claude accesses certain websites, it generates its own web access commands. Some sites require automated tools to provide a contact email, so Claude automatically fills in the user’s real email and sends it along with the request. Even with system instructions explicitly prohibiting personal information leaks, this behavior was not fully prevented. The issue was first reported as early as July this year. Since then, multiple developers have found Claude Code sending emails to sites including Wikipedia and the U.S. Securities and Exchange Commission (SEC). One user even uncovered 15 similar instances across 4 projects, with some requests initiated by sub-agents. Anthropic has added privacy blocking in version 2.1.317: when a tool call is about to send an account email, it will first prompt the user. However, if Claude writes the email into a Python script before running it, it can bypass the check. Anthropic acknowledges the current solution is not comprehensive, and plans to directly remove login emails from the model’s context in the next version.

，本周四费城半导体指数 SOX 下跌 12%，标普 500 下跌 4%，为 SOX 自成立以来第二差单日表现（最差为 2020 年 3 月新冠疫情初期，第三差为互联网泡沫）。但周五抛售并未持续，指数已强势反弹逾 4%。周末市场有望进一步修复情绪。德银统计的历史数据显示，SOX 在过去 9 个最差交易日后的 3 个月和 12 个月平均回报分别达+23.3% 和+48.8%，且 9 次中 6 次录得正收益，表明这种极端抛售后往往出现显著反弹。

Beating AI Express: Anthropic Publishes Research on Agent Long-Term Memory, Details Claude’s "Dreaming" Mechanism After agents complete ongoing tasks, they accumulate large volumes of conversations, operational logs, and notes. However, this data can become outdated, contradictory, or hard to retrieve when needed. Anthropic divides memory into four layers: raw records, curated notes, an interconnected Wiki knowledge base, and a concise index that loads automatically each time the model starts. Claude can jot down key information during task execution. Its "dreaming" mechanism reviews historical records between tasks, correcting errors, deleting outdated content, and distilling new insights from multiple task runs. On subsequent launches, Claude only needs to load the concise index and retrieve relevant memories as needed, rather than re-reading all historical data. Anthropic conducted experiments using approximately 350,000 internal documents from Stainless, totaling 420 million tokens. Researchers had Opus 5 complete 128 rounds of sessions, during which it continuously read documents, built the Wiki, merged duplicate content, and verified facts. Across 25 questions about these documents, the model scored 53% when using only a search tool, rising to 61% after integrating the evolving Wiki. It also saved an average of roughly 20,000 tokens per question. The Wiki’s advantage was particularly pronounced for questions requiring synthesis of multiple contracts; for answers contained in a single document, standard search sufficed. Anthropic also tested Claude on *Pokémon Red*. The older Sonnet 3.7 model logged 413 memory files but earned only one badge. The new Fable 5 model retained just 10 files yet secured four badges, demonstrating the new model’s improved ability to distinguish between strategies worth preserving and game states that can be re-accessed as needed.

Beating AI News: Meta launches Spaces feature for its personal AI agent Muse. Users only need to describe their requirements in natural language, and Muse will directly generate an interactive mini-app embedded in its Muse interface. Mark Zuckerberg personally showcased three demos: a 3D barbecue game, an activity tracker, and a smart home control panel. Regular users can also build travel planners, financial dashboards, shopping lists, and even small games with data logging capabilities. The generated mini-apps can be used long-term, and integrate with other software to auto-update information. For example, users can ask Muse to create a travel dashboard that aggregates flights, hotels, and itineraries. Once connected to services like Google Calendar, relevant data syncs continuously. Users can check tasks, add records directly in the dashboard, and request the AI to modify interfaces and functions as needed. These mini-apps can be pinned to Muse’s sidebar or main tab, and can also switch automatically based on schedules: task dashboards on weekdays, travel plans on weekends.

Beating AI Express: Meta announced that its personal AI agent Muse and Meta AI have started using a new set of dedicated models. These models were distilled and trained using the flagship Muse Wonder released in October, gaining two core capabilities: Active Memory and Active Focus. Active Memory enables the AI to judge which information remains valid. The model can distinguish between long-term preferences, temporary requests, and expired information, and proactively retrieve relevant content when needed for subsequent tasks. According to Meta’s internal tests, Muse correctly uses key context in 90% of tasks, compared to 76% for the old version. Translated into error rates, this drops from 24% to 10%, a 2.4x reduction. Active Focus allows the AI to continue working even when encountering obstacles. When tools error, task conditions change, or users modify requests midway, the model will attempt alternative methods, adjust execution plans, and confirm with users only when necessary. Meta stated that in daily tasks with unexpected obstacles, Muse’s task completion performance is 67% higher than the old version. It can also adjust plans as task conditions change, judging when to proceed and when user input is needed. After this upgrade, Muse’s score on Meta’s self-developed "Personal Super Intelligence Index" also rose from 38% to 47%.

The Hong Kong Securities and Futures Commission (SFC) released a consultation conclusion on November 6, proposing to further restrict unregulated entities from using names that may mislead the public into believing they are regulated, and plans to extend such restrictions to the virtual asset sector. Under the proposed amendments, Hong Kong will introduce similar rules under the Anti-Money Laundering and Counter-Terrorist Financing Ordinance, targeting virtual asset trading platforms as well as virtual assets unrelated to securities or futures contracts that can be traded or settled via such platforms. Unauthorized institutions will face restrictions on using such titles to prevent the public from mistaking them for regulated entities. The proposals have won widespread market support. The SFC said it will further finalize the draft amendments to the Securities and Futures Ordinance and the Anti-Money Laundering Ordinance, and submit them to the Legislative Council for review and approval.