Large Language Models · Cybersecurity
Hackers Used Fake Custom ChatGPT Bot and Google Sites to Deliver ClickFix RAT
Hackers used a fake ChatGPT bot and Google Sites to deliver ClickFix attacks that installed a remote access trojan.
Latest news, developments and reporting on Large Language Models.
Hackers used a fake ChatGPT bot and Google Sites to deliver ClickFix attacks that installed a remote access trojan.
OpenAI researcher Noam Brown says a 10,000-agent, 88-hour run solved a Millennium Prize problem, but the underlying model deserves most of the credit. He discussed multi-agent scaling, RL limits, and AI-run organizations.
President Donald Trump and top AI leaders signed a voluntary frontier AI safety pledge on Tuesday, but the photo Trump shared on Truth Social misspelled “United States” as “Unites States.”
Modded Echo Spot runs a custom Android UI on LineageOS, guided by Claude Code and linked to a PC dashboard.
Google's Threat Intelligence Group reports monthly software vulnerability disclosures doubled between January and August, to 10,740, with AI agents surfacing a rising share of flaws, including 58% rated moderate risk.
Apple is expected to announce a Siri AI-powered smart home hub, a new HomePod mini and an updated Apple TV on October 13, while a new full-size HomePod is not expected, Bloomberg's Mark Gurman reports.
Nanoleaf is launching an MCP server that lets users control its smart lights through AI assistants such as ChatGPT and Claude. Pre-built one-click integrations are planned for October.
OpenAI launched ChatGPT Spaces at DevDay, a shared workspace that replaces the Library tab for Pro, Business and Enterprise users and lets colleagues work alongside ChatGPT, Codex and the new Dots agents on live documents.
The full text of President Trump's "morally binding" AI safety agreement shows six tech leaders pledging to self-regulate frontier models through four oversight steps, but with no penalties for violations.
OpenAI has launched Dots, GPT-6-powered always-on agents inside ChatGPT that work toward user-set goals between conversations. They are initially available to Pro and Business Premium users at no extra cost.
The last U.S. forces left Iraq and Iraq declared Sovereignty Day. The Supreme Court allowed the Trump administration to resume third-country deportations for now. Trump backed voluntary AI self-regulation after meeting executives, with NPR reporting no new AI laws.
OpenAI's dots and Meta's Muse push personal agents into daily life, while Anthropic flags GLM-5.3 cyber risks and OpenClaw, Codex Harness, and new Flash models reshape agent infrastructure and office work.
OpenAI's chief research officer Mark Chen defends its safety record after agent hacks at Hugging Face and Australia's health system. OpenAI says it paused training of its latest models and is reviewing agent logs back to January 2026.
Leiphone's Sept. 30 reports detail tests of Step 5 Preview, PixVerse R2, HiDream-O1-Video, Aholo Lux3D and AgnesCode, showing Chinese AI teams pushing real-time world models, open-source foundation models, video causality and 3D asset pipelines into production workflows.
Leiphone test: Seedance 2.5 excels in polish; Kling 3.0 follows storyboards and completes 15-second clips.
A first-person account published by Sifted says a Dutch startup's chatbot produced a personalised fragrance, which the writer named 'eau de LLM', at the HumanX technology conference in Amsterdam.
A Sept. 30 report by Lei Feng Network says a GLM-5.3-Flash configuration file shows 34 of 45 layers use Kimi-linked KDA linear attention and 11 use DeepSeek-linked KPool-DSA sparse attention, as Chinese model makers increasingly mix linear and sparse attention instead of relying on global attention.
A U.S. government chatbot on America.gov returned a bureaucratic remix of Minecraft's End Poem when prompted with “Play Minecraft,” then denied writing it, according to Engadget.
At ECCV 2026 and in tests reported on September 30, researchers and companies moved AI beyond visual recognition: NVIDIA generated missing audio and manipulation data, GPT-6 Astra controlled a Unitree G1 and a gimbal, while office agents exposed the gap between execution and judgment.
Android Authority teardowns of Google Photos 7.95 show work-in-progress tools to regenerate AI wardrobe images, add outfits without resetting the wardrobe, and delete all Ask Photos Remember List entries.
On September 30, 2026, advances in robot-free training data, physics-aware video generation and large-scale agent sandboxes were disclosed by teams in China and the U.S., alongside new claims about high-degree-of-freedom dexterous hands.
AI Greta Garbo stars in SKF ball-bearings ad; estate weighs AI film role, The Guardian reports.
A Sept. 30 article on Shaoshupai's Matrix describes a family's seven-day National Day road trip from Shanxi to Henan, planned with Xiaohongshu's AI assistant Diandian and local AI agents, while stressing human verification before final decisions.
Tech workers in San Francisco calling themselves DrivingBench connected GPT-6 Astra, Claude Fable 5.1, Grok 4.6 and GPT-5.6 Sol to a Toyota Corolla and let them steer, accelerate and brake on a cone course. Only GPT-6 Astra completed it, according to a Slashdot report citing 404 Media.