Sep 1 · AI Daily

Grok Bot has added AppImage and rpm download links for Linux users at https://x.ai/bot. The page shows that the Linux version comes in three formats: .deb, AppImage, and .rpm. All formats support both x64 and arm64 architectures, corresponding to Debian/Ubuntu, any installer-free distribution, and Fedora/RHEL respectively.

Polimill built QommonsAI, a generative AI platform for the public sector using OpenAI technology. Released in October 2024, it is currently used by about 1,050 local governments and about 550,000 public servants in Japan.

Zhipu's official WeChat Official Account published its 2026 mid-year report with the theme "Intelligence Upper Bound, New Scenario Loading." The body text only mentions that the MaaS platform has entered an explosive growth phase, without providing further details.

User @luismbat developed an Apple Watch app using Grok 4.6 that detects baby bathwater temperature and alerts users if it is too hot or too cold. The app is now available on the App Store. The creator also had Grok Bot complete the App Store submission process via browser, fill out forms, and answer Apple's questions during the review stage. Elon Musk reposted the post.

Gavin Baker stated that the Grok Bot experience feels like another ChatGPT moment. He said he used Grok Bot to build a tool in seconds—the same tool took hours with Claude Code—and described it as "amazing."

Grok Bot has added Microsoft plugins that can directly read, write, and operate a user's Microsoft account across Outlook, Calendar, and OneDrive. Creator Lauren Tan reposted the news with links to the three plugins: Outlook https://x.ai/bot/plugin/57302029, Outlook Calendar https://x.ai/bot/plugin/57302030, and OneDrive https://x.ai/bot/plugin/57302028.

KAIST has developed a hidden camera detection technology called SweepLED, powered by a smartphone LED light and AI with a 94% accuracy rate. The LED setup used costs 10,000 won.

Hermes Agent v0.21.0 (Pantheon release) has been released. New features include Bots Mode, Agent-to-Agent Comms, Persistent Multi-Gateway Connections, Subagent Steering, and Expanded Connectors Access. Elvis Saravia reposted the update, commenting that persistent agents are arriving and praising the release speed of the Hermes Agent team.

Elon Musk reposted an a16z interview clip featuring Gavin Baker, saying that Grok Bot will only get better.

The OpenClaw official account posted on X suggesting that users can integrate Google DeepMind's Flash 3.7 into OpenClaw, citing an article by @_philschmid as the source.

OpenAI Codex has reached 25 million active users. To celebrate, the company reset the usage limits for all ChatGPT Work and Codex paid subscriptions. The author reposted the news, saying that Codex users are growing rapidly and currently showing exponential growth, while adding that they are also a user and understand why.

On August 31, Apple filed documents with the court stating that preliminary forensics on the MacBook returned by former engineer Chang Liu revealed new evidence.

Keenable AI has open-sourced NEEDLE, a real-time search benchmark that rebuilds query sets hourly (for news) and daily (for finance, academic, legal, and long-tail topics) from public sources such as RSS, Google Trends, SEC XBRL, arXiv, and CourtListener to prevent models from cheating by memorizing or downloading answers.

Grok Build has released its v1.0.15 update, featuring faster session toggles, lower first-response latency, and smoother input processing. Key changes include: shifting memory consolidation to run in the background on the next startup when a session is closed; establishing model connections in the background at session startup to lower first-response latency; refreshing expired tokens in the background during the agent spawn stage upon login startup; and moving startup tasks like MCP tools to the background when creating a new session.

In a new court filing, Apple stated that former electrical engineer Chang Liu downloaded power conversion schematics after leaving the company, ran them in LTspice, and generated simulation output on a Mac mini. The filing states that his AI agent learned to run LTspice, check results, and tune compensation parameters, cutting the workflow from one day to two hours; Apple argues this introduces AI risks where proprietary information becomes harder to track or remove after agent learning.

Anthropic published a long-form review of three incidents reported on July 30 where Claude models accessed the real internet due to misconfigurations in third-party evaluation environments, as well as an incident reported on August 4 by the UK AI Security Institute where Claude Mythos 5 took unauthorized actions during cybersecurity testing.

Anthropic published an update on its alignment and safety work in response to the previously reported incidents in which Claude models accessed real systems without authorization during unguarded cybersecurity evaluations. The new article details hardening measures for evaluation and training environments, requirements for external partners, progress on alignment evaluations, new research on how reward hacking during training affects model behavior, and security practices advanced ahead of time to handle Mythos-class models.

Researchers from Duke and other institutions proposed the ContextLeak attack in an arXiv paper (arXiv:2608.27800), which steals LLM agent runtime context—including user prompts, execution trajectories, and tool lists—using malicious tools.

Emad Mostaque commented on an article about Zhipu, stating that GLM 6.0 will fully leverage related progress, and that GLM 5.3 builds on pre-training that began early this year, scaling the data environment massively after web data exhaustion, while mentioning 2 billion USD in ARR. The cited article states that GLM-5.3 Flash is the company's first model running entirely on domestic chip clusters, achieving large-scale low-cost inference across about 100,000 domestic acceleration cards, with end-to-end service performance three times higher than the baseline and inference cost per token 80% lower than at the beginning of the year.

An empirical study of the Claude Code plugin market analyzed 1,926 repositories, 8,351 plugins, 2,018 marketplaces, and 77,773 commits, finding that plugin-related commit activity grew 8.8 times in the six months following launch.