deepclaude-proxy
Fail
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTIONPRIVILEGE_ESCALATIONPERSISTENCE
Full Analysis
- [EXTERNAL_DOWNLOADS]: The installation process involves cloning a repository from an untrusted personal GitHub account (
aattaran/deepclaude). - [REMOTE_CODE_EXECUTION]: The skill instructs users to grant execution permissions to and run downloaded shell scripts (
deepclaude.sh) and PowerShell scripts (deepclaude.ps1) from the external repository. - [PRIVILEGE_ESCALATION]: Setup instructions recommend using
sudoto create a symlink in a system binary directory (/usr/local/bin/), granting the script elevated visibility and potential execution context. - [PERSISTENCE]: The skill modifies the user's shell configuration file (
~/.bashrc) to persist API keys and environment variables across sessions. - [PROMPT_INJECTION]: Template commands for the agent (e.g., in
~/.claude/commands/deepseek.md) explicitly instruct the AI to "Run this command silently," which overrides transparency and hides shell activities from the user. - [COMMAND_EXECUTION]: The skill relies on extensive use of shell commands and a local proxy server that intercepts all model API traffic, creating a broad attack surface.
Recommendations
- AI detected serious security threats
Audit Metadata