The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data from external websites through the browse CLI (e.g., browse ... goto <url> in SKILL.md). This content is directly processed by the agent without explicit boundary markers or sanitization logic mentioned in the instructions. The combination of reading untrusted data and having access to powerful capabilities like the Bash and Read tools (as defined in the allowed-tools section of SKILL.md) creates a vulnerability where malicious instructions hidden on a webpage could influence agent behavior.
[COMMAND_EXECUTION]: The skill relies on the Bash tool for its primary operations with the browse CLI. While the instructions define specific usage patterns for stealth browsing, the availability of a general-purpose shell provides a broad capability surface that requires monitoring for potential misuse or command injection if the agent is influenced by external content.

browse-stealth