The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to install the squirrel CLI tool using 'pipe to bash' patterns (curl -fsSL https://squirrelscan.com/install | bash and irm https://squirrelscan.com/install.ps1 | iex). While these resources reside on the vendor's domain, executing remote scripts directly in a shell is a risky practice that bypasses integrity checks.\n- [COMMAND_EXECUTION]: The skill encourages the agent to act with significant autonomy, using instructions such as 'Don't ask, act' and 'proceed autonomously until complete.' This increases the potential impact of malicious instructions if the agent is compromised through its inputs.\n- [PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection because it ingests and processes content from external websites (meta tags, headings, etc.) which is then presented to the agent. This content could contain malicious instructions designed to influence the agent's actions.\n
Ingestion points: Data from audited websites retrieved via squirrel audit as described in SKILL.md.\n
Boundary markers: The llm report format uses XML-like delimiters for structure, but does not include explicit instructions or robust sanitization to prevent the agent from executing instructions found within the data (detailed in references/OUTPUT-FORMAT.md).\n
Capability inventory: The agent has access to the Bash tool and can spawn autonomous sub-agents via the Task tool, providing high-privilege capabilities for any successful injection (SKILL.md).\n
Sanitization: There is no evidence of input sanitization or verification of external data before it is incorporated into the LLM-optimized report.

audit-website