# AI safety
Latest news and articles about AI safety
Total: 16 articles found
Alibaba’s ‘Wukong’ Aims to Turn DingTalk into a Corporate AI Engine — Safe, Embedded and Built for B2B
Alibaba has launched Wukong, an enterprise‑native AI platform embedded into DingTalk, designed to automate workflows while enforcing strict data access and audit controls. The initiative, placed under a new Alibaba Token Hub led by CEO Wu Yongming, signals a strategic pivot toward B2B AI where safety, permissions and skills integration become the primary commercial levers.
China’s Yushu CEO Says the ‘ChatGPT Moment’ for Embodied Robots Is Near — But Not Here Yet
At the Yabuli forum Yushu Technology CEO Wang Xingxing defined a practical threshold for an embodied-AI “ChatGPT moment” and said it may take two to three years to achieve. He emphasized that improved motion capabilities are the essential prerequisite for robots to perform real-world tasks and that progress will come through parallel advances in hardware and software.
China’s New ‘GEO’ Economy: Firms Paying to Seed and ‘Poison’ AI Recommendations
Chinese media exposed a growing industry—known as GEO—that creates and distributes coordinated promotional content to bias AI models’ outputs in favour of paying clients. By automating content production and leveraging networks of publishing accounts, firms can cause mainstream models to recommend fabricated or promoted products, posing risks to consumer trust and market fairness.
Small Beijing Firm Behind 'GEO' AI Tool Named on China's 315 Show, Raising Data‑poisoning and Supply‑chain Alarms
China's 315 consumer‑rights broadcast named a product called the "LiQing GEO optimization system", linking it to a small Beijing company with limited staff and modest capital. The appearance of this vendor in a national probe underscores risks from opaque suppliers in AI data and model supply chains and points to mounting regulatory and market pressure for greater transparency and provenance controls.
OpenAI Pauses Promised 'Adult Mode' to Focus on Core AI Improvements Amid Competition and Oversight Concerns
OpenAI has delayed its planned ChatGPT "adult mode" to prioritize core product improvements such as model intelligence and personalization. The move comes amid fierce competition and internal debate over the company’s ethics, including a resignation tied to a contentious U.S. Department of Defense partnership and revisions to that contract to limit surveillance and weaponization uses.
OpenAI Tweaks ChatGPT’s ‘Instant’ Brain with GPT‑5.3 — And Promises GPT‑5.4 Even Sooner
OpenAI has launched GPT‑5.3 Instant, an update designed to make ChatGPT’s most used mode more helpful by reducing unnecessary refusals and improving context comprehension. The company also signalled a faster timetable for GPT‑5.4, suggesting more substantial capabilities — and more intense scrutiny — lie ahead.
Beijing Opens Data and AI Security Testing Centre to Anchor Safe AI Growth
Beijing has opened a municipal Data and AI Security Testing Centre to provide testing, risk assessment and standards work for AI systems, based in MenTouGou’s Jingxi Zhigu industrial cluster. The centre aims to professionalise AI assurance, supporting the capital’s digital-economy goals while raising compliance requirements for developers and shaping regional standards.
Anthropic Backs Away: Safety Pledge Softened as Competition and Policy Uncertainty Bite
Anthropic has watered down its 2023 Responsible Scaling Policy, dropping a blanket pledge to pause model scaling when safety cannot be proven and replacing it with conditional delays tied to competitive position. The change reflects commercial pressures, a fragmented U.S. regulatory landscape and an intensifying race among leading AI developers.
Musk Opens Grok 4.2 Candidate to Public Beta, Promising Weekly ‘Fast‑Learning’ Updates
Elon Musk has opened a candidate public beta of Grok 4.2, requiring users to opt in and inviting public feedback. The model claims a new fast‑learning capability and will receive weekly updates accompanied by release notes, accelerating xAI’s iterative development approach but raising questions about safety and oversight.
OpenAI Recruits Creator of OpenClaw, Vows to Keep Viral Agent Open-Source via New Foundation
OpenAI has hired Peter Steinberger, creator of the widely adopted agent framework OpenClaw, and pledged to place the project into a foundation that will keep it open-source and independent while receiving funding and support. The move is a tactical win for OpenAI but raises questions about governance, security and the balance between openness and centralization as agent platforms mature.
US Military Allegedly Used Anthropic’s Claude in Venezuela Operation, Raising Questions About AI’s Role in War
U.S. media report that Anthropic’s AI model Claude was used in the January 3 U.S. operation in Venezuela, routed via a partnership with Palantir. Anthropic has not confirmed the claim and stresses its policy forbidding uses that facilitate violence, but the allegation raises legal, ethical and strategic questions about private AI models in military operations.
OpenAI’s Voice Models Tapped for Pentagon Drone‑Swarm Challenge, Raising Dual‑Use Concerns
OpenAI has shared an open‑source voice‑to‑instruction model with two Pentagon‑selected defence firms competing in a prize to produce voice‑controlled drone‑swarm prototypes. The move highlights the tension between commercial AI innovation and the risks of rapid diffusion of components that can enable more autonomous and potentially weaponised systems.