OpenClaw, an open-source AI assistant project, released version v2026.5.27 today. The update promotes OpenAI-compatible vector embedding services — both local and hosted — to a built-in core engine, retiring the old plugin-based compatibility registration. It also integrates Pixverse video generation, with API region selection and external plugin packaging, natively filling the AI video creation gap in its agent ecosystem. Additionally, the new version unlocks vLLM deep thinking parameter integration and allows direct routing of bare Anthropic model identifiers.
On the security front, OpenClaw tightens isolation. Group prompt text is now completely separated from system prompts to prevent malicious injection, and the update blocks command wrappers with side effects and unsafe Node runtime environment variable overrides. To prevent unauthorized Tailscale network exposure, node and device role approval permissions are now restricted to administrators. For instant messaging, Telegram message sending has been switched to a persistent delivery queue, improving success rates across network fluctuations.