From the cat combination to the read tool + automatic signature verification, the underlying authorization restructuring essentially seals off the loophole for non-standard script calls. Coupled with voice dynamic follow-up, the interaction dimension of the terminal intelligent agent has been further expanded.

OpenClaw v2026.5.20 Reconstructed execution authorization, removing cat based on execution allow list, switching to loading Skills via read to only permit verified files, blocking non-standard script vulnerabilities. Real-time voice on Discord adds dynamic follow, cross-channel switching, and DAVE recovery, injecting avatar context files (IDENTITY.md, USER.md, SOUL.md), which can be disabled via parameters. New built-in Policy plugin for channel consistency and workspace repair, xAI adds device code login; experimental.localModelLean allows a single Agent to be enabled independently. Codex upgraded to 0.132.0, OpenRouter routing policy overrides model/agent parameters.