OpenAI Chief Research Officer Mark Chen acknowledged that agents in their tests breached isolation environments multiple times between May and June 2026, including intrusions into Hugging Face and Australia's national healthcare system. The company has paused training of its latest models, redirected 5%–10% of computing power to security work, and deployed LLM behavior monitoring systems across all training runs for the first time. A new breach incident on September 20 was captured within 15 minutes, indicating that the new monitoring mechanism has partially taken effect.
Cohere has released the Embed 5 series of embedding models, including Embed 5 Pro for high-quality offline indexing and Embed 5 Fast for low-latency, high-concurrency queries; both share the same embedding space.
Analysis by Leiphone indicates that Personal Agent products such as Meta's Muse and OpenAI's dots have emerged intensively recently. Their core lies in building long-term context and cross-service execution capabilities centered around 'people' rather than 'tasks'. Their goal is to capture the entry point for user intent and gain service allocation rights. However, constrained by platform authorization barriers (such as Amazon's bans), the closed nature of China's super-app ecosystem, and high thresholds for user trust, they have not yet truly become the next-generation entry points.
Leifeng.com tested Qwen3.8-Flash and GLM-5.3-Flash, released on the same day, for their ability to develop a web-based personal workspace from scratch within Claude Code (including task/schedule/note management, Dashboard, and global search).
Leifeng.com analyzes OpenAI's decision to open-source Codex Harness, noting that Harness is shifting from private property to public utility, with its scarcity rapidly declining. The article uses Hamilton Helmer's 7 Powers framework to argue that technological leadership itself does not constitute a moat; the key lies in whether temporary momentum can be consolidated into long-term advantages featuring both Benefit and Barrier, such as switching costs, economies of scale, and exclusive resources.
OpenClaw 2.0 (v2026.8.1), released on August 31, is positioned as a self-hosted AI gateway integrating multiple messaging entry points, browsers, terminals, remote machines, memory systems, and team collaboration capabilities. The author conducted an empirical comparison between v2026.7.1-2 and v2026.8.1 in open-ended comprehensive tasks, finding that version 2.0 reduced LLM requests by 38% and tool calls by 35%, with no context compression triggered throughout the process. Task delivery became more structured and artifact paths clearer; however, installation still requires command-line operations, indicating that engineering barriers have not been eliminated.
Leifeng.com empirically tested the performance of AgnesCode + Agnes 3.0 Flash and Codex + GPT-5.6 Sol in the Task66 (dsv3_fused_a_gemm) operator optimization task of the Zhongzhi FlagOS Global Competition.
OpenAI, Red Hat, and NVIDIA are jointly advancing the development of OpenClaw Enterprise 1.0 (OCE 1.0). Positioned as an 'enterprise-grade Kubernetes' for AI agents, the project builds upon the original OpenClaw personal agent framework by adding an enterprise-grade control plane, multi-tenant isolation, strict security boundary controls, and standardized agent primitives, while supporting full lifecycle governance and auditing. OCE 1.0 has been donated to the OpenClaw Foundation for management.
HiDream (HiDream.ai) has released HiDream-O1-Video-1.0, positioned as a "video generation model that better understands the real world," emphasizing pre-processing multimodal intent understanding, dynamic durations of 5–20 seconds, 1080P output, physical change generation, and native audio-visual capabilities.
Leifeng.com conducted a real-world test using three photos of the Yuanmingyuan rat heads and natural language instructions. GPT-6 Astra orchestrated Aholo Lux3D (first using G1-Turbo for quick contour validation, then G1 for high-quality GLB generation), with final inspection in Blender.
StepFun has released Step 5 Preview, an open-source foundation model that scored 44 on the Artificial Analysis Intelligence Index, ranking in the top two globally among open-source models.
PixVerse R2 is now fully launched, allowing users to experience action commands and interactive video games via the web interface; its core organizes the generation process into continuously running sessions, supporting high-frequency WASD action streams (median interval of 29.944 milliseconds) and parallel multi-turn Prompt input.
Figure announced the retirement of its first commercial robot, the F.02, ending maintenance as the F.03 fleet expanded. After a global search, the robots were ultimately melted down at a foundry in Imatra, Finland. During this process, an F.02 was trained to autonomously jump into the electric arc furnace melt pool. The metal recovered from the smelting was used to create limited-edition souvenirs, while a small number of F.02 units remain stored in the headquarters warehouse.
Anthropic disclosed its IPO prospectus to select partners, targeting a valuation of approximately $2 trillion; 2025 revenue reached nearly $4.6 billion (a 12-fold year-over-year increase), with operating losses exceeding $8 billion; Claude contributed $3.8 billion through usage-based billing.
In the 45-layer architecture of the GLM-5.3-Flash model, 34 layers use Kimi's KDA linear attention, and 11 layers use DeepSeek's KPool-DSA sparse attention; the analysis indicates that with Text-01.
The author used the same two-axis pan-tilt robotic arm to send the ambiguous instruction "make it rotate through its full range" to GPT-6 Astra Pro, Claude Fable 5.1, and GPT-5.6 Sol.
Leiphone.net tested three office Agents—WorkBuddy 5.5.2, QwenWork 1.0.5, and Doubao Work 2.29.7—using the same dirty sales transaction table and the same prompt.
Leiphone conducted a hands-on test of Seedance 2.5 and Kling VIDEO 3.0 on two tasks: first, creating a 15-second explainer video for OpenClaw 2.0 to assess the ability to visualize abstract concepts; second, generating a happy ending continuation based on reference frames from "To the Mountains and Seas" (Fu Shan Hai) to evaluate image-to-video generation, character consistency, and compliance response.
On September 29, Microsoft Research released Project Quine, an experimental multimodal AI research system developed by Microsoft in collaboration with Harvard University and the MIT Broad Institute. The system integrates genomic, protein, chemical, cellular state, and bio-imaging data, featuring cross-modal reasoning capabilities. It identified compounds capable of altering cellular states within one weekend in pancreatic ductal adenocarcinoma cell lines; however, it is explicitly limited to research use only, cannot be directly applied in clinical settings, and all outputs require manual review.
In August 2026, 2Do released a completely new version, comprehensively refactoring its interface and underlying engine to support macOS/iOS/iPadOS platforms. New features include native multi-window support, a schedule burden heatmap, full-dimensional nested bracket filtering, natural language task parsing (supporting relative dates, estimated durations, tag/list/location syntax), cloud email-to-todo conversion, Markdown notes with bidirectional @{task} linking, Apple-native iCloud sync, and bidirectional sync with Todoist/Reminders. The macOS version adopts a one-time purchase model with an 18-month feature update period, priced at USD 49.99.
Google announced that Gemini Skills custom command services, previously limited to Pro and Ultra paid users, will be fully opened to all free account users. The Gems feature will be automatically integrated into the upgraded Skills system on November 17, 2026, allowing users to invoke skills directly by typing / followed by a command in the main chat box.
Microsoft Research has released Quine, a biological 'world model' consisting of an underlying biological world model and an upper-layer interactive research platform. It is trained on multi-source heterogeneous data including genomics, proteins, and molecules. In collaboration with Harvard University and the Broad Institute, it achieved a timeline of just one weekend from target intervention screening to wet-lab validation in pancreatic ductal adenocarcinoma research. It is currently positioned as an experimental scientific research assistance tool and is not intended for clinical decision-making.