@manateelazycat
manateelazycat says, in Chinese, that GLM's security research ability is amazing: after they submitted the omarchy-desktop-top plugin to the Omarchy plugin store yesterday, Omarchy's security bot said the plugin had a security problem; Codex said an array out-of-bounds issue was indeed possible, but when asked to research the code it said "This content can't be shown" on security grounds, and they then asked the offline GLM 5.3 Flash, which spent about 5 minutes on the code, confirmed the vulnerability, said it was more serious than the GitHub issue described, gave a complete solution and a fix patch, and ran automated security attack-and-defense tests, linking the marketplace issue comment and the fix commit; they say that even when only fixing security vulnerabilities in local software, not attacking networks, offline GLM 5.3 is much stronger than GPT: not that GPT lacks the ability, but that its online safety policy refuses normal requests; the still is a dark terminal agent session with "Ran" steps, an "Explored" block, a "Failed (exit 1)" python3 step, and a "This content can’t be shown" notice reading "We take extra care with some cybersecurity requests."