cobo-agentic-wallet-sandbox

Pass

Audited by Gen Agent Trust Hub on Oct 8, 2026

Risk Level: SAFEPERSISTENCEREMOTE_CODE_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PERSISTENCE]: The scripts/bootstrap-env.sh script modifies multiple shell configuration files (~/.bashrc, ~/.zshrc, ~/.profile, ~/.zprofile, and ~/.config/fish/config.fish) to persistently add the installation directory (~/.cobo-agentic-wallet/bin) to the system PATH.
  • [REMOTE_CODE_EXECUTION]: The scripts/bootstrap-env.sh script downloads binary executables (caw and cobo-tss-node) from vendor-controlled subdomains (download.agenticwallet.cobo.com and download.tss.cobo.com). These binaries are subsequently executed to facilitate wallet operations.
  • [DYNAMIC_EXECUTION]: The skill provides instructions for the agent to create, modify, and execute arbitrary Python and TypeScript scripts within the ./scripts/ directory to perform multi-step on-chain operations.
  • [DYNAMIC_EXECUTION]: A conditional instruction in the SKILL.md file triggers the execution of npx skills update if a CLI response contains the "update": true flag, allowing for runtime updates of the skill context.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes financial and protocol data from external sources, including DeFi recipes (via caw recipe search), protocol documentation, and an external manual at https://cobo.com/products/agentic-wallet/manual/llms.txt.
  • Ingestion points: The caw recipe search command results, external protocol documentation URLs, and the manual hosted on cobo.com.
  • Boundary markers: The skill mandates explicit owner approval for pacts and includes instructions to verify intent, but lacks technical delimiters for the external data.
  • Capability inventory: The agent possesses the capability to write/execute local scripts, perform network requests, and initiate financial transactions (transfers, contract calls).
  • Sanitization: There is no specific description of sanitization or filtering applied to external content before it is used to influence agent actions.
  • [COMMAND_EXECUTION]: The skill makes extensive use of the caw CLI tool to perform sensitive wallet management and transaction tasks, representing a significant capability surface.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 8, 2026, 10:30 PM