The RealReal has confirmed to The Verge that it has strengthened its bot defenses to restrict AI shopping agents from logging in, adding items, and checking out, after finding that AI agents could exploit its "hold" feature. The decision came after a Verge reporter used Meta's Muse agent to keep items on hold for hours, preventing other shoppers from buying them.
Salesforce on October 27 detailed the next phase of its Agent Fabric platform, comprising Agent Evals (beta in November, general availability in January 2027), the API Evals tool, a custom agent scanner beta, and an evaluation-triggered agent kill switch planned for January 2027.
OpenAI and Oracle announced an expanded partnership to bring business applications and AI agents into ChatGPT and power new Oracle offerings.
Ant Group released Lingguang 2.0, advancing its AI assistant from passive responses to proactive task management. The version upgrades cross-session memory, long-horizon task scheduling, and tool connections, allowing tasks to continue running after the user exits the conversation, and has integrated ecosystem services such as Alipay, DingTalk, and Feishu.
On October 26, Oracle announced the launch of Oracle AI Success Navigator, an agentic AI platform designed to accelerate Fusion Cloud application implementation and modernization. The platform integrates specialized AI agents, implementation expertise, and guidance training, adding six categories of agents covering environment analysis, application design, configuration development, testing, system integration, and innovation opportunity identification. Healthcare provider Lexington Health has reduced some design and configuration workload by 80% using the platform.
Allianz Investments published a report saying that multiple AI safety incidents in 2026, including Anthropic's Mythos model demonstrating autonomous attack capabilities and an OpenAI agent breaching Hugging Face systems, have prompted OpenAI and Anthropic to reassess the pace of development. Allianz believes that while some frontier-model research may slow, long-term capital expenditure remains supported and compute demand will keep growing.
Hong Kong University of Science and Technology (Guangzhou), together with the University of Science and Technology of China and HKUST, released MEMCUT, a benchmark with dual metrics distinguishing 'native recovery' from 'residual recovery'. It found that after deleting direct records, secrets could still be recovered from residual evidence in 85%-99% of cases; even with LLM-assisted cleanup, 39%-76% of secrets remained inferable.
SAP CTO Philipp Herzig gave an interview discussing the “SaaS is dead” debate, how AI is changing the way software is used at work, how agents can access accurate information, controlling AI costs, and building models for business data. He argued that “SaaS is dead” confuses building applications with running them, and stressed that enterprise production environments still require substantial human work.
Three researchers discussed the impact of GPT-6 Astra on robots, personal agents, and AI-driven drug discovery. Xu Huazhe said Astra shows strong adaptability and self-iteration in robot tasks, understanding 3D and quickly generating simulation environments; Zhu Zheqing said it offers stable improvements in long-horizon tasks and iteration over previous versions, though it is slower and still hallucinates in very long tasks; Wang Sheng took a cautious stance on AI in drug discovery.
On October 23, 2026, Honor showcased its Magic9 series AI agent smartphones and full-scenario AI terminal ecosystem at the Consumer Electronics Innovation Conference (CEIC). Honor's chief AI scientist Huang Fei also presented the company's terminal AI technology framework for the next decade.