The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches its core instructions and logic from an external domain (https://ethskills.com/audit/SKILL.md) at runtime. This creates a dependency on remote content that is not versioned or controlled within the skill package itself, allowing for silent updates to the skill's behavior.
[REMOTE_CODE_EXECUTION]: The installation instructions in catalog.json direct users to run claude plugin install https://github.com/austintgriffith/ethskills. Installing plugins from an unverified third-party GitHub repository allows for the execution of arbitrary code within the agent's environment, which can lead to full system compromise.
[DATA_EXFILTRATION]: The skill's stated purpose includes 'filing GitHub issues' based on synthesized findings. This entails sending code analysis results to external servers and requires the agent to have write access to GitHub repositories, providing a mechanism for unauthorized data transmission.
[COMMAND_EXECUTION]: The setup and demo sections rely on shell commands such as curl to fetch remote resources and plugin managers to install external code, which are executed directly in the user's environment.
[PROMPT_INJECTION]: The skill processes untrusted external data in the form of smart contract code. It lacks sanitization or boundary markers for these inputs and involves a complex multi-agent synthesis chain. This creates a significant surface for indirect prompt injection, where malicious instructions hidden in a contract's comments could manipulate the auditing agents' results or actions (Category 8 surface).

Smart Contract Audit