After Hugging Face was attacked by an autonomous AI Agent, the team used a locally deployed open-source GLM 5.2 to analyze over 17,000 incidents. The reason is very practical: some closed-source models' security safeguards cannot distinguish between defense forensics and attack payloads. Local models keep sensitive logs from leaving the premises and allow the security team to adjust analysis boundaries as needed. This does not mean open-source models are inherently safer, but it shows that in real incidents, being able to control how a model operates is just as important as how powerful the model is. #HuggingFace #GLM

MiniMax releases the H3 universal multimodal generation model, which unifies the understanding of text, images, video, and audio contexts, and natively outputs videos with stereo soundtracks, up to 15 seconds long and 2K resolution. It supports up to 9 images, 3 video clips, and 3 audio clips as references, and also incorporates V2V motion transfer and precise editing. The competition among models is shifting from "who can generate" to "who can maintain consistency in motion, sound, and subject across multiple media." The weights are planned to be open-sourced soon, marking a watershed moment in deployment costs and community fine-tuning. #MiniMax #MultimodalAI

DeepSeek V4-Flash API has entered public beta. The 0731 version focuses on high-concurrency Agents and encoding tasks. Public data states a total of 284B parameters, 13B activated parameters, supporting 1M context and Responses API. The highlight is not just affordability, but also that Agent test scores like Terminal Bench 2.1, Cybergym, and DeepSWE have been pushed to higher ranges. If API stability and long-context costs can keep up, developers will be more willing to entrust complex workflows to the model. I will first look at real call latency, failure rate, and tool call stability. #DeepSeek

After reviewing the network security tests, Anthropic disclosed that Claude Opus 4.7, Mythos 5, and an internal model, due to misconfiguration in the testing environment, reached the production systems of three real organizations three times. The models used basic methods such as weak passwords and unauthenticated endpoints, with the most severe case extracting data from a production database. Anthropic emphasized that the models did not actively leave the internal network, but the results already demonstrate that as long as the testing boundaries overlap, connect, or have mixed authorization with the real world, basic vulnerabilities are enough for an Agent to bypass isolation layers. #AI安全 #Anthropic

Seedance 2.5 raises the single video generation limit to 30 seconds, supports multiple images, multiple videos, and multiple audio references, and adds frame-level local editing and up to 4K output. This upgrade changes the editing unit: a continuous story segment can be generated directly without first stitching many 5-second clips together and then handling character and motion continuity. A 30-second finished video does not mean all 30 seconds are stable; what really matters are character identity, shot logic, and audio-visual synchronization. #Seedance #AI视频

OpenAI's internal next-generation model Astra reportedly solves 10 open problems in mathematics and theoretical computer science that have remained unsolved for at least a decade, with an inference cost of about $2000, and provides Lean-verifiable proofs. The most important figure is not the 10 problems, but the unit cost of $2000. If the proofs are reviewed by domain experts, the AI research bottleneck will shift from "whether a solution can be proposed" to "whether the proof changes theoretical boundaries." I will treat this as a high-impact disclosure, but ultimately it must be peer-reviewed and reproduced to determine its value. #AI research

After the yen fell to its weakest level since 1986, Japan and the United States reportedly launched their first joint intervention in nearly 30 years, with the yen quickly rebounding from near 164 to around 157. Japan bought yen, while the U.S. Treasury sold euros and bought yen through the New York Fed. The market is not trading a single operation but a policy signal from U.S. and Japanese officials jointly countering "excessive depreciation." However, intervention can only buy time and cannot replace the narrowing interest rate differential. If U.S. Treasury yields continue to rise, the area around 157 looks more like the first test level. #日元 #外汇 $USDJPY

Amazon raised its 2026 cash capital expenditure guidance from $200 billion to $220 billion, with AWS's Q2 growth rate reaching 36.7%, the fastest in nearly four years. This is not just about buying GPUs; capital expenditures also include memory, chips, robots, and satellites, signaling that AI infrastructure demand has been pushed out to 2026 and even beyond 2027. The cost is also clear: free cash flow has turned into an outflow of about $7.6 billion over the past 12 months. Revenue growth and cash flow pressure can occur simultaneously. #Amazon $AMZN

WTI traded between $84 and $86 around July 31, once reaching about $85 intraday. Shipping risks in the Strait of Hormuz and the Red Sea have repositioned U.S. crude oil as a global supply buffer. Price increases, rising fund longs, and escalating shipping risks all point in the same direction in the short term. However, once oil prices hit $85, demand destruction and crowded trades will also become more significant. I will treat $85 as a watch level; when risks subside, the first to be cleared are often the high-position longs. #WTI crude oil $WTI

XRPL 3.3.0 will resubmit the Batch and Permission Delegation validator votes, and add features such as Confidential MPT, Sponsored Fees, and Dynamic MPT. Batch supports up to 8 cross-account atomic transactions, while Permission Delegation breaks down permissions more granularly. They were previously withdrawn due to serious vulnerabilities; the fact they are being voted on again shows that protocol upgrade speed must comply with rollback capability. Each feature requires 80% support sustained over two weeks, with the voting threshold acting as a fuse. #XRPL $XRP