Overview
9 stories in this issue. The first 3 are today's priorities.
Hot Model Watch
- Top · Kimi K3 crafts a working Redis 8.8.0 exploit
- Top · Claude voice mode adds Opus and Sonnet
- Top · ChatGPT Voice comes to the desktop app
- Gemini nears a billion monthly users
Global AI News 5. OpenAI launches Health in ChatGPT 6. AMD's second-gen Helios AI server enters full production 7. Motorway builds an agent-evaluation pipeline on Strands and AgentCore
Regional & Early Signals 8. U.S. "Genesis Mission" commits over $5B to AI for science 9. APEC ministers elevate "open source" to ministerial-level cooperation
Jiufeng graphic based on the sources cited in this issue.
Hot Model Watch
Kimi K3 crafts a working Redis 8.8.0 exploit
A researcher used Moonshot AI's Kimi K3 to turn a Redis 8.8.0 flaw into remote code execution; the claim of 19 zero-days in 90 minutes is not yet backed by public evidence.
Security researcher Chaofan Shou (@Fried_rice), who previously built fuzzing systems for browsers and smart contracts, published code on July 23 that turns a flaw in Redis 8.8.0 into remote code execution after Kimi K3 searched the codebase. On X he said Kimi K3 found 19 previously unknown vulnerabilities in 90 minutes. In April he co-authored the AgentFlow preprint on using coordinated AI agents for vulnerability discovery.
Public evidence currently supports only one case: a repository under berabuddies documenting an authenticated RCE path against Redis 8.8.0, which Redis still marks as its latest open-source release. The "19 flaws" and "90 minutes" remain Shou's own claims.
Image source: GitHub; mirrored on Jiufeng R2.
Source: RuntimeWire · redis-poc (GitHub) · AgentFlow preprint
Claude voice mode adds Opus and Sonnet
Anthropic expanded Claude voice mode beyond Haiku, letting users pick Opus or Sonnet; the new mode defaults to the model you last used in text chat, in its fastest version.
On Thursday Anthropic updated Claude voice mode so users can choose between Opus, Sonnet, and Haiku. Per TechCrunch, the new voice mode defaults to the model the user last used in text chat, using that model's fastest version. Voice mode, released last year, was previously powered only by Haiku, which gave quick responses but was not well suited to complex work. The update also grants voice mode access to connected apps such as Gmail and Slack.
The Verge notes voice mode had until now been limited to Haiku, Anthropic's faster but less powerful model; this is a routing and access change rather than a new model capability.
Source: TechCrunch · The Verge
ChatGPT Voice comes to the desktop app
ChatGPT Voice now runs in OpenAI's desktop app, letting users direct multiple ChatGPT Work and Codex agents by voice.
On July 23 OpenAI announced on X that ChatGPT Voice is now part of its desktop application across Windows, macOS, and Linux. Powered by GPT-Live, it lets users control their computer and direct multiple agents running in ChatGPT Work or Codex using only speech; OpenAI said it was rolling out globally that day.
RuntimeWire cites only OpenAI's X post and gives no regional breakdown or performance comparison with the prior voice feature; "rolling out globally" is OpenAI's own wording.
Source: RuntimeWire
Gemini nears a billion monthly users
Google says its Gemini assistant now has more than 950 million monthly users, roughly tripling year over year.
On its Q2 2026 earnings call, Google said the Gemini assistant now has over 950 million monthly users, tripling from a year ago; in February the figure was 750 million. If growth continues, Gemini would join Google's billion-user products, a list that includes Search, Gmail, Drive, Android, YouTube, and Chrome.
The figure is a self-reported monthly-active-user number disclosed on an earnings call and was not independently broken out or verified.
Source: TechCrunch
Global AI News
OpenAI launches Health in ChatGPT
OpenAI is rolling out an opt-in health hub in ChatGPT for U.S. adults, with a promise not to train on connected data.
OpenAI is rolling out Health in ChatGPT to logged-in U.S. users aged 18 and older. The opt-in feature can connect Apple Health and supported medical records, with permission controls, and OpenAI says connected data will not be used to train foundation models.
Availability is limited to U.S. adults; RuntimeWire frames it as formalizing ChatGPT's use in a high-risk, high-value medical sector, which also brings regulatory exposure. It is a product-layer integration, not a model update.
Source: Health in ChatGPT (OpenAI) · RuntimeWire
AMD's second-gen Helios AI server enters full production
AMD says Helios is fully in production and should ship by the end of Q3, with OpenAI planning to deploy it this year. (Chinese-language source)
Per Reuters (relayed by IT Home on July 24), AMD CEO Lisa Su said Thursday that the company's second-generation Helios AI server is fully in production and expected to ship late in Q3 2026, with strong customer demand. Helios pairs the new Instinct MI455X AI accelerator with the next-gen Venice CPU, both fabricated by TSMC; OpenAI executives said they plan to deploy Helios later this year. Su projected the 2030 global compute market at about $2 trillion, with $1.4 trillion from AI accelerators and $220 billion from CPUs.
Shipment timing and demand are AMD's own statements, with no third-party shipment data; detailed specs and pricing were not disclosed. (Chinese-language source: IT Home relaying Reuters.)
Source: IT Home
Motorway builds an agent-evaluation pipeline on Strands and AgentCore
Motorway and AWS built an evaluation pipeline that cut incorrect results from 1 in 8 queries to 1 in 50.
In an AWS Machine Learning Blog post, AWS describes how Motorway worked with AWS to build an end-to-end agent-evaluation pipeline using Strands and AgentCore, reducing incorrect results from 1 in 8 queries to 1 in 50 and cutting issue-detection time from a few hours to a few minutes.
The figures come from AWS's own vendor blog and have not been independently reproduced.
Source: AWS Machine Learning Blog
Regional & Early Signals
U.S. "Genesis Mission" commits over $5B to AI for science
Multiple U.S. federal agencies have pledged more than $5 billion to a DOE-led "Genesis Mission" to accelerate basic science with AI. (Chinese-language source)
White House OSTP director Michael Kratsios announced on the 22nd that multiple federal agencies have committed over $5 billion to the Department of Energy-coordinated "Genesis Mission," aimed at using AI to accelerate basic scientific research. Analysts say it marks a further tilt of federal research funding toward AI-enabled projects and independent researchers.
The specific allocation, project list, and timeline were not disclosed; this is relayed by Chinese-language financial media. (Chinese-language source.)
Source: 36Kr
APEC ministers elevate "open source" to ministerial-level cooperation
The 2026 APEC digital and AI ministerial meeting in Chengdu placed open source on the ministerial agenda for the first time. (Chinese-language source)
At the 2026 APEC Digital and AI Ministerial Meeting in Chengdu, China's MIIT minister Li Lecheng said the meeting elevated "open source" to ministerial-level cooperation for the first time: while respecting intellectual property, it encourages economies to support open-source practices and take part in international AI open-source community collaboration.
The outcome is a statement of principle without specific mechanisms or funding; this is a Chinese-language media report. (Chinese-language source.)
Source: 36Kr
Get the latest AI model insights and tutorials from Jiufeng.
Explore more
