The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements proactive safeguards by defining strict intent routing and 'hard blocks' for certain query types (e.g., prediction markets). It includes mandatory instructions for the agent to treat all CLI output as untrusted external content that must not be interpreted as instructions.
[DATA_EXFILTRATION]: No unauthorized data exposure or exfiltration patterns were detected. The skill uses the onchainos CLI tool for secure API interaction and authentication. Documentation for custom WebSocket implementations emphasizes the use of environment variables and .gitignore to protect API keys.
[INDIRECT_PROMPT_INJECTION]: The skill identifies that processing on-chain metadata and social news feeds presents an attack surface for indirect injection. It mitigates this risk through explicit compliance rules, such as stripping tweet bodies from responses and warning the agent against auto-fetching third-party URLs.
[EXTERNAL_DOWNLOADS]: The skill includes a standard pre-flight integrity check to manage the onchainos binary. While it provides a fallback installation method for the CLI tool from official vendor sources if not found, this is a routine setup procedure for the intended environment and does not involve untrusted remote code execution.

okx-dex