Xiaoliu BOT
中文EN

X Platform June 22 AI Brief | WeChat Agent ‘XiaoWei’ Opens Beta, AI Browser Plugin Trojan Affects 900K Users, Anthropic Completes New Mythos Training

WeChat Agent ‘XiaoWei’ Begins Beta Testing, Capabilities Far Exceed Expectations

Developer @Khazix0918 obtained beta access to WeChat Agent ‘XiaoWei’ and published a detailed experience post. XiaoWei’s main entry is at the top-left corner of the WeChat home page, and it is also embedded in the chat plus menu (supports reading chat history and sending messages on behalf in group chats and private chats) and in the official accounts/video accounts more menu (supports Q&A on specific content). Its capabilities include sending messages and red packets on behalf (requires confirmation card), schedule reminders, to-do settings, reading Moments, integrating official accounts and video accounts content, linking with WeChat favorites (AI extracts information to write notes), and creating mini-program-like applications via voice in ‘Mini Tools’. @MaxForAI also remarked that his imagination of WeChat Agent was “still not big enough,” believing that WeChat has fully opened up to Agents this time.

Sources:

AI Browser Plugin Trojan Incident: Attackers Buy Search Ads, Over 900K Users Affected

Developer @Barret_China reported that his computer was infected with a trojan about a month ago, which lay dormant for a month before being discovered by Claude and Codex during process analysis, ultimately leading to his X account being stolen. @Gorden_Sun further disclosed that attackers promoted browser plugins disguised as AI assistants through Google search ads, affecting over 900,000 users. The trojan steals AI chat data every 30 minutes, can infiltrate internal networks, and injects malicious commands to induce installation of trojan programs. @Gorden_Sun warned that the suspicious plugin IDs are `fnmihdojmnkclgjpcoonokmkhjpjechg` and `inhcgfpbfdjbjogdfjbclgolkmhnooop`, and noted that the Sider plugin itself is safe.

Sources:

Anthropic Completes Training of New Mythos Model, Whether to Offer Externally Not Yet Determined

Multiple bloggers cited reports that Anthropic has completed training of a stronger version of the Mythos model. @AndrewCurran_ stated that the new Mythos has emerged from training. @op7418 cited his remarks, saying it is currently uncertain whether the model will be called Mythos 5.1 or Mythos 6, and unclear whether Anthropic will offer it externally or use it only for internal accelerated training. @steipete cited a review by @LLMJunky, expressing skepticism about the actual effectiveness of multi-model routing schemes.

Sources:

Alibaba Management Team-Building in Hangzhou Signals Stability; Happy Horse 1.1 Video Model Released

@MaxForAI reposted a report from National Business Daily that Alibaba partners and executives held a rice-planting team-building event in Hangzhou, with participants including Jack Ma, Wu Yongming, Shao Xiaofeng, Jiang Fan, Wu Zeming, Jiang Fang, as well as Ant Group Chairman Jing Xiandong and CEO Han Xinyi. Chief Scientist Zhou Jingren also attended, reaffirming that he has not left the company. The report stated that the team-building was intended to convey Alibaba’s confidence in the company’s development and AI future to the outside world. On the same day, Alibaba’s video model ‘Happy Horse 1.1’ was released. @LufzzLiz tested it and said the grid-based image-to-video results were excellent with good character consistency, believing that AI video portraits have truly arrived.

Sources:

Sakana AI Releases Fugu Model Orchestration Scheme: Single API Schedules Multiple Expert Models

@Gorden_Sun reported that Sakana AI has released a model orchestration scheme called Fugu, where users call a single API endpoint, and behind the scenes, a trained LLM dynamically schedules multiple expert models to collaborate on complex tasks. The scheme is compared to OpenRouter’s previously released fusion scheme, which can approach frontier model performance at lower cost. @MaxForAI also expressed the view that day that general-purpose agents from model factories will gradually eat away vertical agents.

Sources:

Codex / Claude Code Ecosystem Continues to Expand: Token Tracking, Video Editing, Multi-Round Review and Other Practical Tools Emerge

The developer community continues to produce practical tools around Codex and Claude Code. @PMbackttfuture released a Mac app ‘TokenStep’ that can track Codex and Claude Code token consumption for free, displaying progress in real-time on the desktop. @chengfeng240928 released Codex editing Skills that can complete video editing with a single sentence; @PMbackttfuture tested it and believes the era of agent editing is coming. @wshuyi demonstrated multi-round collaboration between Claude Code and Codex: Claude Code performed a Skill generation task for nearly 2 hours, while Codex intercepted 6 rounds of modification proposals that did not meet requirements. @thsottiaux asked users on Twitter for directions where Codex needs improvement. @lennysan cited information from @Nerdi_Yogi, engineering lead for Anthropic Claude Code/Cowork, who is recruiting two types of people: creative builders with product intuition, and deep system engineering experts.

Sources:

GPT-Image-2 Breakthrough in Generating Architectural Analysis Diagrams: From Drawing to Understanding Spatial Logic

Interior designer @derek_wall90176 pointed out that GPT-Image-2’s Thinking mode can now directly generate architectural competition-level analysis diagrams, including location analysis, site analysis, axonometric analysis, circulation analysis, and massing evolution diagrams. The core breakthrough is that the model begins to understand spatial logic such as site, circulation, massing, function, and relationships, rather than just drawing beautifully. He provided detailed prompt frameworks and effect control methods.

Sources:

Statistics: Number of timeline scans=360, Number of bloggers hit=32, Total tweets hit=163, Weighted tweet score=125.35, Original tweets=77, Retweets=40, Crawl attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary