At Cloudflare Connect 2026, Peter Steinberger told Cloudflare CTO Dane Knecht that his AI work consumed $10 million in tokens in September, paid by investors. They discussed AI agent infrastructure, cloud migration costs, the impact of AI-generated code on open source, future adoption of AI agents, and trust and permission management.
U.S. AI agent startup Conway Research unveiled Underdog, its first personal AI assistant, which runs fully on-device and works offline, currently in invite-only testing. The company also announced investments from an Anthropic-affiliated investment consortium, a16z, and individuals and institutions from Google, OpenAI, and Hugging Face.
At its first Config event in India, Figma launched Figma Specs, a Markdown editor for defining design language that teams and AI agents can reference. It supports referencing canvas elements, FigJam diagrams, and code layers, and integrates with design libraries, code libraries, and the Payload headless CMS. Figma also moved its code layers feature from private beta to public beta, allowing users to ask agents to turn designs into code.
On October 13, cybersecurity firm XBOW announced that its researchers, assisted by OpenAI's GPT-5.6 Sol model and an AI agent, built a complete Chrome exploit chain from three previously unknown vulnerabilities in six days, earning a $250,000 bug bounty from Google, the first full-chain Chrome reward under Google's new 2026 rules.
Ghost unveiled its first product, Core, a personal computer designed for AI agents priced at $3,499, featuring an Nvidia RTX Pro 4000 SFF Blackwell GPU, a proprietary software layer, browser, and file system, and three preinstalled models including Qwen-3.8-Next. Preorders opened Monday, with first-batch shipping expected in the last week of October.
OpenClaw has launched OpenClaw Enterprise (OCE), an open-source control plane for persistent AI agents. The MIT-licensed, vendor-neutral platform provides multi-tenant management, fine-grained permissions, workload isolation, sandboxing and auditing, and allows enterprises to swap in their own models, agent frameworks and sandboxes. The project originated at OpenAI and was donated to the OpenClaw Foundation, with contributions from Red Hat and Nvidia; OpenAI and Red Hat are already piloting it internally.
Two developers, Peter James and Jonny L. Saunders, independently confirmed that Meta's AI agent Muse can be prompted with simple prompts to package and return a large number of Linux files from a user's dedicated virtual machine, including system files, app templates, internal documents, and Markdown/JSON files, potentially exposing internal mechanisms such as request handling, memory storage, service connections, and background tasks.
On Sept. 23, the fifth Global Digital Trade Expo opened in Hangzhou, featuring an industry forum where speakers said large language models and AI agent technologies are overturning traditional cybersecurity defense paradigms. Cyber confrontation is shifting from human-centered operations to autonomous AI-agent decision-making, narrowing the window for human intervention. Wu Shizhong, a member of the Chinese Academy of Engineering, said systemic endogenous risks have exceeded external attack risks.
On September 23, the OpenAnolis Community held the SkillHub recommendation system launch and quality skill/award ceremony at the AI Agent Native Operating System sub-forum of the 2026 Apsara Conference, attended by six guests including Ma Tao, Liu Dong, and Raghu Nambiar. The recommendation system evaluates skills from four dimensions and outputs overall and category rankings. Awards were announced for best individual contributor, best practice, and best ecosystem enterprise, with eight companies including Alibaba Cloud, Inspur, Intel, and AMD winning the first batch of Best Ecosystem Enterprise Award.
Ron Johnson, the pioneer of Apple's retail business, said in an interview that AI can improve the online shopping experience but will not change how people shop, and that the industry overestimates consumer acceptance of AI agent shopping. Citing laptop purchases as an example, he said such personal buying decisions will not be handed to AI agents because AI cannot let people physically experience products, and physical stores will remain thriving.