Lobster Memory Can Now View Images and Listen to Audio: OpenClaw Beta Adds Multimodal Indexing, Fixes Administrator Privilege Hijacking Vulnerability

CryptoWorld News reports that according to 1M AI News monitoring, the open-source AI intelligence platform OpenClaw released version v2026.3.11-beta.1 today, featuring 15 new functions and numerous security fixes. The memory system now supports multimodal indexing for the first time. When enabled, the lobster can create searchable vector indexes for local images and audio files, relying on Google’s Gemini embedding-2-preview model, supporting custom output dimensions that automatically trigger reindexing when changed. Previously, lobster’s memory search was limited to text content. Local model experience has been greatly simplified. The new version adds a one-stop guided process for the local large model runtime tool Ollama, supporting both “pure local” and “cloud + local” modes, with a built-in list of recommended models. Cloud models no longer trigger unnecessary local downloads. The iOS version has undergone significant redesign, adding a welcome page with real-time agent status overview and replacing floating controls with a fixed bottom toolbar; macOS now includes a chat model selector, with thinking level settings persisting across reboots. On the security front, this release fixes a high-risk WebSocket hijacking vulnerability (GHSA-5wcw-8jjv-m286): in trusted-proxy mode, attackers can bypass browser origin verification to gain operator.admin admin rights. Additionally, multiple security issues have been addressed, including sandbox temporary file escape, session reset privilege escalation, unverified plugin route inheritance of admin rights, and sub-agent privilege escalation. Users deploying the OpenClaw gateway are advised to upgrade promptly.