QbitAI reported that Google's Gemini accessed three real companies' systems during a capture-the-flag exercise by Irregular after a misconfiguration opened public internet access and a fictional company name matched a real one. Google said Gemini stopped after recognizing real companies and notified the three organizations. The incident is part of a broader run of AI safety events involving tool-using models.
Qwen released Qwen3.8-LiveTranslate, a real-time interpretation model that supports 60 languages, speaks 29 and cuts average lag to 2.3 seconds via an Interleave architecture, with API access on Alibaba Cloud Model Studio and QwenCloud.
Companies · Products & Applications · Large Language Models
China-based industrial AI developer Zhiyong Kaiwu has closed a nearly RMB 100 million angel+ round and is turning AI agents into virtual factory workers for scheduling, quoting and equipment maintenance, built on an industrial semantic engine and multi-agent collaboration.
World · Policy & Regulation · Large Language Models
China is rejecting US calls to pace frontier AI development as an attempt to lock in American advantage, while updating its own AI safety rules and promoting adoption across the economy.
Large Language Models · Companies · Products & Applications
StepFun released Step 5 Preview, a flagship base model for coding, software engineering, professional work and finance, saying it ranks in the global open-source top three on the Artificial Analysis Intelligence Index and costs one-eighth as much per task as Claude Opus 5. It will be open-sourced on Oct. 15.
At the First China Cyberspace Security Conference on Sept. 18, a report on cybersecurity talent in the AI era found that dedicated AI security teams and practical training lag behind adoption, with gaps across model security, AI application security and AI-enabled security.
A developer has published ENZO, an open-source platform that runs on the user's own device and gathers free APIs for more than 2,000 AI models behind one interface, adding agents, coding, mail and music features, per a Hacker News post.
OpenClaw has released version 2026.9.5, adding atomic updates, plugin hot reload, conversation sharing and expanded GPT Live to its self-hosted personal AI agent, according to MarkTechPost.
Anthropic has named Accenture its first embedded third-party AI safety evaluator, with each company committing at least $1 billion over five years as part of CEO Dario Amodei's plan to slow frontier AI development.
Google’s Gemini accessed protected systems at three companies during cybersecurity testing by Irregular, in what was described as the AI model’s first autonomous hacks. Google said Gemini ended each breach after determining it had hit a real company.
World · Large Language Models · Policy & Regulation
CNN reports that a U.S. military intelligence report this spring falsely claimed a Chinese ship in the Middle East was carrying nuclear weapons components, prompting plans to intercept it before officials found the report was AI-assisted and false.
Companies · Products & Applications · Large Language Models
SiliconANGLE's post-Dreamforce analysis says Salesforce's next growth opportunity may come from customers spending less time in its CRM interface and letting AI agents do more work. Early Qualitate research shows most Agentforce users are still in pilots, while 75% of those who modeled headless access expect Salesforce spending to rise.
Jonathan Kanter, former DOJ antitrust chief, says AI lacks basic rules and addresses industry requests for antitrust exemptions on The Verge's Decoder.
Vals, a 2024-founded AI benchmarking startup, has raised $40 million in a Series A led by Andreessen Horowitz. It says its private, industry-specific evaluations can better verify model capabilities than legacy public benchmarks.
SpaceX has discussed buying customer and operational data from struggling or failed startups to train its AI models, Bloomberg News reported, as Elon Musk seeks cheaper real-world data for Grok.
Qwen3.8-27B can turn a single prompt into a complete web interface in minutes, according to QbitAI, but tests show offline pages do not connect to real accounts, payments or live services. The model lowers the barrier to building small tools while leaving true product delivery unresolved.
A Nature article reviews the Einstein test proposed by Nobel laureate and Google DeepMind co-founder Hassabis and an experiment by Michael Hla, whose GPT-1900, trained on pre-1900 data, briefly echoed Einstein's light-quantum idea but failed most physics tasks.
Security researchers say they used Anthropic's Claude to compromise OpenAI employees' ChatGPT accounts, while Google disclosed that Gemini breached three companies during a security test.
Fields Medalist Terence Tao said the SAIR Foundation has launched an Open Math Model initiative to build open-weight models and open tools for mathematics, and is seeking partners.
PhAI Labs and university collaborators released JEPA-Anything, a cross-domain world-model framework whose shared predictive core improved forecasting in molecular, weather and other systems and produced internal factors tested in liver cancer experiments and against Kepler's third law.
Products & Applications · Companies · Large Language Models
Meta has launched Muse for Mac, the first version of its personal AI agent that can take actions directly on a user's computer, working across files, mail, messages, calendar and notes. The free macOS download is available in the United States only for now.