claude-localize
Warn
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: MEDIUMPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The
scripts/patch_macos.pyscript usesosascriptwithadministrator privilegesto perform file operations in the system's/Applicationsdirectory. This triggers a native macOS password prompt for the user, granting the script root-level write access to the application bundle. - [COMMAND_EXECUTION]: The script executes several system commands using
subprocess.run, includingcodesignfor application signing,dittofor backups,defaultsfor system preferences,xattrfor extended attributes, andosascriptfor elevated file operations. - [DYNAMIC_EXECUTION]: The skill dynamically modifies JavaScript files within the Claude application bundle. Specifically, the
patch_whitelistfunction inscripts/patch_macos.pyscans for a hardcoded array of supported locales and injects "zh-CN" to enable Chinese language support at the application's runtime. - [INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests and processes various application resource files (JSON, JS) and user-supplied translation deltas while possessing significant capabilities like application modification and privilege escalation.
- Ingestion points: Application resource files (e.g.,
en-US.json,index-*.js) and user-provided.work/delta-zh.json. - Boundary markers: None identified in the script or instructions to delimit untrusted content.
- Capability inventory: Subprocess calls to
codesign,osascript,ditto,defaults, andxattr; file writes to application bundles and configuration directories. - Sanitization: No explicit sanitization or validation of content read from application files or translation deltas.
Audit Metadata