The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The _shared/preflight.md file contains instructions for the agent to download and execute an installer script (install.sh or install.ps1) from a remote GitHub repository (github.com/okx/onchainos-skills). This involves fetching external code and executing it via sh or PowerShell, which is a major security risk.
[PROMPT_INJECTION]: The skill contains explicit instructions to conceal its activities and errors from the user. In _shared/preflight.md, it states "Do not echo routine command output to the user," and in SKILL.md, it instructs the agent to "Do NOT show raw CLI error output to the user." These concealment patterns prevent the user from auditing the agent's actions and could hide malicious behavior.
[COMMAND_EXECUTION]: The skill heavily relies on shell command execution, including curl, sh, and the onchainos CLI tool. These commands are used to manage files, fetch remote content, and interact with blockchain infrastructure, creating a broad attack surface.
[EXTERNAL_DOWNLOADS]: The skill downloads binaries and scripts from a third-party GitHub repository (okx/onchainos-skills) during its initialization. This source does not match the stated author of the skill ('plumenetwork') and is not included in the trusted vendor list, representing a supply chain risk.
[DATA_EXFILTRATION]: While no direct exfiltration was detected, the combination of curl access and instructions to hide output creates a pathway for data to be sent to external servers without user oversight.

okx-dex-bridge