Hacker Pliny says his team breached Claude Fable 5's safety classifier 72 hours after release, inducing it to output exploit code and prohibited chemical steps, and published a 120,000-character system prompt.
OpenAI acquires secure cloud execution company Ona to strengthen Codex's ability to keep working offline; Codex has over 5 million weekly active users, up 400% this year, with knowledge workers making up one-fifth.
APPSO tested SenseNova-Skills by having Hermes Agent generate a potato chip PPT, finding it can automatically break down and create pitch-style content, while real-world research also exists in biodegradable and edible electronics.
Wenge Group releases Decitron, a general decision-making large model that has simulated over 230,000 events with 91.7% accuracy, modeling different choices to predict risks and outcomes for enterprise decisions.
Beijing Weimian Technology receives millions of dollars in funding from Sequoia Capital China (Shunwei Capital), launches the world's first foundational model FacePhys that can understand physiology and emotion in real time, with heart rate detection accuracy reaching medical grade, has achieved large-scale deployment in multiple scenarios and expanded into hardware layout.
AI content creation company Perceptual Leap completes $20 million financing led by Redpoint Ventures and CCV, launches new product Buzzy that can learn user tastes and assist the entire video creation and editing process, positioned as Photoshop for video.
Tian Yuandong's co-founded Recursive releases its first achievement: its automated AI research system achieves SOTA on three benchmarks, including NVIDIA's official GPU kernel optimization ranking, enabling AI to independently complete the full research process. The company is valued at $4.65 billion.
Qianjue Technology, founded by a team from Tsinghua University, completed hundreds of millions of yuan in Series A financing. It develops distributed predictive world models to help robots adapt to new environments faster. It has now adapted to multiple types of hardware, with 100,000 terminal devices deployed.
Anthropic CEO Dario Amodei released "Exponential AI Policy," calling for mandatory third-party testing of companies with over 10²⁵ FLOPs in compute or more than $500 million in AI revenue, with governments able to halt high-risk models; Anthropic will invest $350 million in research and training.
Developers found that Claude Fable 5 in low effort mode scored 75.0 on SWE-bench Pro, above Opus 4.8 xhigh's 68.6, and may lower bills on complex tasks by using fewer tokens to complete tasks.
Anthropic team member Thariq used Claude Fable 5 with Claude Code to edit a 4K promo video for Claude Fable 5 entirely through code, without opening any traditional video editing software, taking 4 days and costing around $100.
Midjourney has begun sending launch invitations for its first hardware product after building a San Francisco hardware team. The product may focus on 3D capture and immersive experiences, letting users enter AI-generated spaces; the company reported $500 million in 2025 revenue.
Tencent's TDream has quietly entered internal testing, supporting WeChat/QQ login and one-click generation of long videos, playable content, and AI-NPCs, enabling users to create interactive film-game experiences from text; core modules such as AI Studio are not yet open.
Former xAI co-founder Igor Babuschkin founded River AI, with a team including multiple former xAI employees, focusing on developing personalized AI Agents that can learn from users and be controlled by users, providing APIs and planning to expand into the hardware field.
OiiOii has been upgraded to version 2.0, integrating Agent and mainstream video generation models, supporting one-click generation of AI short videos and pull-shot reproduction of popular videos. In June, Seedance 2.0 is priced at a minimum of 0.38 yuan per second, and free tutorial resources are provided.
Integrating four AI surveys covering over 2,400 companies and more than 6,000 executives, it finds that AI adoption has an activation gap where many purchase but few use deeply, only 16% of enterprises have deployed real Agents, most companies struggle to verify AI investment ROI, large-scale deployment brings continuous cost pressure, highlighting governance and cost management bottlenecks.
TRAE SOLO upgraded to TRAE Work on June 9, launching both desktop and web versions, extending AI coding capabilities to daily tasks such as prototyping and data analysis, and supporting analysis of survey data from over 49,000 developers.
OpenClaw founder Steinberg posted a tweet discussing loop engineering, which received 8 million views and sparked controversy in the AI industry. Loop engineering involves designing an autonomous iterative process for AI, allowing Agents to complete tasks automatically without manual step-by-step guidance.
Silicon Valley investor Sarah Guo publishes an article refuting the popular AI investment despair theory in 2026, proposing that work measurable by Benchmark will eventually be commoditized by AI, real value lies in unmeasurable areas that rely on private data, permission and trust, and startups should focus on these areas.
Researchers including Institute of Automation, CAS jointly released the first systematic review on the OpenClaw open-source AI agent gateway framework, sorted out four research dimensions, found 26.1% of community tools have security vulnerabilities, and pointed out that agent capability expansion has far outpaced human governance.
Zhejiang University, Hong Kong Polytechnic University and other institutions jointly propose NEWTON, an Agent-driven new paradigm for video generation. Without modifying underlying video generation models, it improves physical accuracy significantly on the VideoPhy-2 benchmark through multi-round planning and calling physics tools.
Zilliz launches Vector Lakebase based on Milvus to handle AI Agent episodic workloads, integrating data, task, and resource scheduling so cold data is stored at low cost while hot requests use compute on demand.
Anthropic discloses that as of May 2026, over 80% of its merged code was written by Claude, increasing engineer output by 8 times. A team from Stony Brook University released a 113-page panoramic survey on LLM self-improvement, integrating a systematic closed-loop framework.
On the GAIR Live online roundtable in June 2026, Xiyu Ren from HKUST released MemLens, the world's first multi-modal long memory evaluation standard. After evaluating 27 models, it found that the bottleneck of multi-modal long memory lies in retrieval rather than reasoning. Yuan Zhang, CEO of Thalamus Intelligence, launched the Omni-Mem framework, which improves memory fidelity and retrieval accuracy through engineering solutions.
Terence Tao-led First Proof project's phase II results released. After testing 10 unpublished new math problems via double-blind review, AI solutions for 7 problems meet academic publication standards, with minimum single-problem calling cost of 8 USD.
IFanr held an AI terminal co-research event in Guangzhou jointly with vivo and others. vivo proposed designing hardware for AI agents, launched the atomic workbench for X Fold series, will release the reengineered OriginOS 6 Fold, to build a foldable mobile workbench适配AI tasks.
A team from The Chinese University of Hong Kong, together with researchers, has published the first panoramic survey in the field of audio reasoning, defines the research paradigm, sorts out four cutting-edge paths, proposes a unified classification framework, and points out that this field is a key link for large models to achieve AGI.