autoglm-browser-agent
Fail
Audited by Gen Agent Trust Hub on May 19, 2026
Risk Level: HIGHCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs users to execute
xattr -d com.apple.quarantineon its bundled binaries (relayandmcp_server). This command explicitly removes macOS security protections designed to prevent the execution of unsigned or untrusted software. - [REMOTE_CODE_EXECUTION]: The skill distributes and executes pre-compiled binaries in the
dist/directory. These opaque executables constitute a risk as their internal logic is not auditable within the skill source code. - [EXTERNAL_DOWNLOADS]: Mandatory dependencies include the
mcporterNPM package and a specific Chrome extension (AutoClaw, ID:jelniggicmclhfgnlapbkgfibmgelfnp) required for the automation server to function. - [DATA_EXFILTRATION]: The skill includes functionality to capture and transmit browser screenshots to external IM services like Feishu (Lark) via API, creating a path for exfiltrating sensitive visual information.
- [PROMPT_INJECTION]: Employs aggressive instructional overrides, using terms like '๐จ ๆ้่ฆ็่งๅ' (Most important rule) and '็ปๅฏนไธๅ ่ฎธ' (Absolutely not allowed), to force specific agent outputs and suppress the agent's default summarization or safety behaviors.
Recommendations
- AI detected serious security threats
Audit Metadata