The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content, creating a vulnerability to indirect prompt injection. An attacker could embed malicious instructions in the HTML or structured data of a website being audited, which might then be interpreted by the agent after it processes the audit findings.
Ingestion points: scripts/aeo_audit.py fetches HTML from external URLs and reads local files.
Boundary markers: No structural delimiters or explicit instructions are provided to the agent to isolate and treat audited content as data only.
Capability inventory: The skill includes scripts for network interaction (urllib.request) and file access, and it references external command-line tools like ghost-cli.
Sanitization: The HTML parsing logic in scripts/aeo_audit.py does not include sanitization or safety checks for embedded instructions.
[DATA_EXFILTRATION]: The script scripts/aeo_audit.py performs network requests using urllib.request.urlopen. While this functionality is intended for fetching website content for SEO auditing, it represents a network capability that could be used to communicate with arbitrary external servers.
[COMMAND_EXECUTION]: The skill provides local Python scripts (scripts/aeo_audit.py, scripts/build_prompt_matrix.py) and references external CLI tools like ghost-cli for SEO tasks. This encourages the agent to execute shell commands, which increases the attack surface if combined with untrusted data processing.

seo