Lobster Memory now supports image-text fusion: OpenClaw beta adds multimodal indexing, fixes admin privilege escalation vulnerability

According to 1M AI News monitoring, the open-source AI agent platform OpenClaw today released version v2026.3.11-beta.1 of its beta, which includes 15 new features and numerous security fixes.

For the first time, the memory system supports multimodal indexing. Once users enable it, the lobster can build a searchable vector index of local image and audio files, relying on the Google Gemini embedding-2-preview embedding model at its core. It supports custom output dimensions and automatically triggers reindexing when the dimensions are changed. Previously, the lobster's memory search was limited to text content.

The local model experience has been significantly simplified. The new version adds a one-stop guided flow to the local large model running tool, Ollama, supporting both "purely local" and "cloud + local" modes. It includes a built-in recommended model list, and cloud models no longer trigger unnecessary local downloads. The iOS version has undergone a major redesign, adding a welcome page with a real-time AI agent status overview and replacing floating controls with a bottom-fixed toolbar; the macOS version adds a chat model selector, and the thought level setting can persist across restarts.

On the security front, this release fixes a critical WebSocket hijacking vulnerability (GHSA-5wcw-8jjv-m286): in trusted-proxy mode, attackers could bypass browser referer validation to gain operator.admin administrator privileges. Additionally, it addresses several security issues, including sandbox temporary file escape, session reset privilege escalation, unauthenticated plugin route inheriting admin permissions, and sub-agent privilege escalation. Users deploying the OpenClaw gateway are advised to upgrade promptly.