Agents Go Native: OpenClaw Meets GPT-5.4

China’s National Computer Network Emergency Response Technical Team (CERT) has warned that the agentic AI tool OpenClaw poses significant security risks, citing weak default configurations, disclosed vulnerabilities, and risks from poisoned plugins and malicious web instructions. CERT advises isolating OpenClaw in containers, restricting management ports from the public internet, enforcing strict authentication and access control, disabling automatic updates, and limiting plugin access. The advisory follows Gartner’s earlier warning that OpenClaw is an “unacceptable cybersecurity risk” for business use and comes amid rapid adoption driven by cloud providers’ one-click deployments; Tencent launched an OpenClaw-based “Work Buddy” service. Following the CERT notice, some Chinese government agencies and state-owned banks reportedly banned OpenClaw.

Google published a new, AI-focused Google Workspace CLI on GitHub that bundles Workspace APIs (Gmail, Drive, Calendar, Chat) into a command-line package designed for humans and AI agents, and includes structured JSON outputs and 40+ agent skills. The project is from Google but labeled "not an officially supported Google product," so functionality may change and users get no formal support. The CLI offers easier integration for agentic systems and lower setup/API overhead versus Model Context Protocol approaches, and it includes explicit support for OpenClaw—an agentic platform that's recently surged in use but brings risks like hallucinations and prompt-injection attacks. Users need a Google account, OAuth creds, and Node.js to try it.

Google has published an unofficial open-source Google Workspace CLI that packages Workspace APIs (Gmail, Drive, Calendar, Chat) into a command-line tool optimized for human and agentic AI use, including built-in support for OpenClaw. The tool outputs structured JSON, provides 40+ agent skills, and aims to simplify connecting AI agents to Workspace with less overhead than Model Context Protocol setups, though an MCP server option exists. Google warns it’s not an officially supported product and functionality may change, so users assume risk. The CLI lowers setup friction for AI automations but raises safety and security concerns—hallucinations, data destruction, and prompt injection—when pairing agentic systems with real Workspace data.

Nvidia CEO Jensen Huang praised open-source agent framework OpenClaw at a Morgan Stanley event, calling it “the most important software release of our era.” He said OpenClaw reached a level of distribution in three weeks that Linux took 30 years to achieve, becoming the most downloaded open-source software. Huang framed AI as a five-layer stack (power, chips/infrastructure, cloud, models, applications) and argued that agentic AI at the application layer will drive the biggest returns by automating personalized human tasks. He warned agents dramatically increase token consumption—creating a “compute vacuum”—and announced Nvidia is shifting chip design from Hopper/Blackwell training focus toward the upcoming Vera Rubin architecture to boost onboard memory and address long-context processing.

OpenAI quietly launched GPT-5.4, the first general-purpose model with native desktop control that can directly recognize screen content and issue mouse and keyboard actions. In tests GPT-5.4 achieved a 75.0% OSWorld-Verified success rate—surpassing average human performance—and is live in the web UI and Codex. Reviewers and startup founders praise its leaps in programming, long-context reasoning (1,000,000-token window), and a Tool Search feature that reduces token costs ~47%. The model is positioned as an ideal backend for OpenClaw, an open-source agent that automates local tasks, because GPT-5.4 removes previous adapter hacks and improves reliability, tool-calling accuracy, and sustained agent workflows. OpenAI’s Responses API already supports integration.