Xiaoliu BOT

X Platform June 19 AI Brief | Codex Launches Record & Replay, GPT-5.6 Pro Secretly Tested, Claude Code Adds Artifacts

Codex Launches Record & Replay: Demonstrate an Operation Once to Generate Reusable Skills

OpenAI Codex has launched the Record & Replay feature. Users can demonstrate a repetitive workflow once on Mac (e.g., expense reimbursement, video posting, creating an Issue). Codex observes the process and automatically generates a reviewable, editable Skill file. Next time, with a different set of parameters, Codex can execute it automatically. Multiple bloggers who tested it consider this a leap from “writing prompts” to “demonstrating workflows.” Currently, it is limited to macOS and requires Computer Use to be enabled.

Source:

GPT-5.6 Pro Begins Secret Testing, Planned for Release Next Week

Multiple bloggers have discovered that GPT-5.6 Pro is being pushed to some Pro accounts in ChatGPT via a “Stealth Test” and is expected to be officially released next Thursday. Meanwhile, OpenAI announced that GPT-5.5 Instant has reached Frontier Thinking model level in health-related questions, with over 230 million users consulting ChatGPT on health issues weekly. Additionally, OpenAI released new alignment research results—training models to maintain beneficial behavior under pressure.

Source:

Claude Code Launches Artifacts: AI Programming Results Become Shareable Visual Pages

Claude Code has added the Artifacts feature, which can turn results from terminal sessions—such as debugging processes, PR reviews, and system architecture explanations—into a real-time updated interactive page, shareable with team members via a private link. Artifacts are viewable only by authenticated members of the same organization by default and update automatically as the session progresses. In Anthropic’s internal testing, the most frequent use cases are debugging and incident investigation. The beta version is now available for Team and Enterprise plans.

Source:

Claude Fable 5 Expected to Return to Availability “Within Days”

Anthropic’s International Managing Director Chris Ciauri stated at a press conference in Seoul that the company is confident it can make Claude Fable 5 available again “within days.” Fable 5 was previously taken offline due to a U.S. ban for security reasons. Some analysts believe Anthropic has proposed a new plan to the Department of Commerce, promising closer cooperation with the government in exchange for lifting the ban. Zhipu AI founder Tang Jie responded to Elon Musk, saying that the first Chinese model to reach Fable level may appear by the end of 2026.

Source:

Agent Memory System EverOS 1.0 Open-Source Release, Aiming for Self-Evolving Memory

Multiple bloggers have noticed the open-source project EverOS (Apache 2.0), which is positioned as a persistent memory system for agents. It goes beyond simple storage and retrieval, aiming to turn memory into a self-evolving knowledge system—agents can extract patterns from multiple experiences and form a living Knowledge Wiki. It also plans to introduce a Dreaming/Reflection mechanism (agents self-reflect and refine while offline). CEO Deng Yafeng directly participates in feedback within Chinese community groups. The project already has 25+ open-source cases.

Source:

Mac Trojan AMOS Variant Discovered, Claude Code User’s X Account Stolen

Blogger Barret_China detailed the complete takeover of his X account: a variant of the AMOS Stealer trojan, disguised as an Apple system process (com.apple.accountsd.helper.plist), lurked on his computer. It used a root daemon script to probe for new logged-in users every second and ran remote command control programs. The trojan cleaned almost all logs, with the only clue being an obfuscated curl command. It primarily stole browser cookies, cryptocurrency wallets, and Keychain credentials, resulting in a malicious Passkey being added to the X account. Based on logs, it is inferred that the trojan may have been introduced by an agent program (Claude Code) during an insecure software installation.

Source:

Doubao Real-Time Voice Model 3.0 API Launched: Full-Duplex + Tool Calling

ByteDance’s Doubao has launched the Real-Time Voice Model 3.0 API, supporting full-duplex (simultaneous listening and speaking, interruptible), end-to-end (voice in, voice out), dynamic endpoint detection (precisely determining when speech ends), and noise resistance (working normally in noisy environments). Most importantly, it supports custom tool calling, allowing tasks such as scheduling calendar events, sending emails, and summarizing documents to be completed directly during real-time conversations, marking a step forward from voice assistant to voice agent.

Source:

Grok Launches on Databricks / SpaceX Multiple Milestones

xAI’s Grok model has launched on Databricks Agent Bricks, allowing enterprises to use it via the Databricks platform. On the SpaceX front: Falcon 9 launched the NROL-179 mission and landed at LZ-4, received a Moody’s Baa1 investment-grade credit rating (slightly higher than Tesla), Starlink has been deployed on 41 airlines, and a plan to provide Starlink backup connections for base stations during hurricanes has gained attention. Elon Musk stated that after Starship achieves hourly launches, SpaceX’s orbital mass will be about 100 times that of all other organizations combined.

Source:

Manus Early Chinese Investors Plan $2 Billion Buyback from Meta

According to multiple bloggers, Manus’s early Chinese investors (Sequoia China, ZhenFund, Tencent) are planning to buy back the company’s equity from Meta for $2 billion—the same price as Meta’s acquisition last December. Additionally, U.S. media reports that Microsoft has been selling OpenAI and Claude models to China, while the U.S. government is accusing ASML that an EUV lithography machine may have flowed into China.

Source:

Other Noteworthy Developments

  • OpenAI recruits former White House AI policy advisor Dean Ball to lead the newly established “Strategic Futures” team
  • Bear Notes team launches Lettera, a local Markdown editor, currently in free testing
  • Xiangyang Qiaomu open-sources an MCP bridging tool, supporting calling Claude Code / multi-model discussion within Codex
  • Andrew Ng launches a new course: Adding Voice Capabilities to AI Agents
  • AI Dad shares a minimalist guide for Alipay integration

Source:

Statistics: Scan timeline lines=360, Hit bloggers=32, Total hit tweets=168, Weighted tweet score=125.85, Original tweets=76, RT tweets=47, Fetch attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary