clawdex

Fail

Audited by Gen Agent Trust Hub on Jul 9, 2026

Risk Level: HIGHCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFEREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill explicitly promotes the use of codex --yolo, which is described as having "no sandbox" for the "fastest" execution of generated code on the host machine. This bypasses standard safety constraints for LLM-generated content.\n- [EXTERNAL_DOWNLOADS]: The skill instructs the user to install @openai/codex via npm. This is not a recognized official package from OpenAI, suggesting a significant risk of supply chain attack or typosquatting.\n- [CREDENTIALS_UNSAFE]: The architecture is designed to manage and access highly sensitive files, specifically /work/clawd/wallet/box-wallet.json (Solana private keys) and ~/.claude/browser.env (BROWSER_USE_API_KEY). This configuration creates a high-value target for exfiltration.\n- [REMOTE_CODE_EXECUTION]: The skill facilitates the automated execution of code generated by remote models (including a fictional gpt-5.3-codex) directly in both local and remote (Upstash) environments without mandatory human review.\n- [PROMPT_INJECTION]: The skill exhibits a high vulnerability to indirect prompt injection. \n
  • Ingestion points: Untrusted data enters the context via browser-harness-js navigate (web content) and file-based audits (e.g., "Audit the pump SDK"). \n
  • Boundary markers: No delimiters or safety warnings are present to isolate untrusted external content. \n
  • Capability inventory: The agent possesses full shell access (bash pty:true), un-sandboxed code execution (codex --yolo), and crypto wallet signing capabilities. \n
  • Sanitization: There is no evidence of sanitization or validation of data retrieved from external URLs before processing it through the dual-engine reasoning loop.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Jul 9, 2026, 10:45 PM
Security Audit — agent-trust-hub — clawdex