The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates browser automation, which involves interacting with and ingesting untrusted content from external websites. \n
Ingestion points: The skill uses agent-browser open to visit URLs and snapshot -i to read the page structure. \n
Boundary markers: There are no explicit instructions to the agent to ignore or delimit potentially malicious instructions embedded in the target website's content. \n
Capability inventory: The skill can perform actions like click, fill, and eval on the ingested pages, providing an exploitation surface. \n
Sanitization: The instructions do not include steps to sanitize or filter the data retrieved from web pages before the agent interacts with it. \n- [DYNAMIC_EXECUTION]: The skill utilizes dynamic script execution to handle complex browser interactions that standard commands cannot address. \n
Evidence: The skill recommends using agent-browser -p kernel eval for executing custom JavaScript on the page. \n
Evidence: For cross-origin iframes, the skill suggests using kernel browsers playwright execute to run Playwright scripts against the session. \n- [COMMAND_EXECUTION]: The skill uses shell commands and pipes to manage browser sessions and extract connection details. \n
Evidence: SKILL.md provides bash snippets that use kernel browsers list piped into jq to resolve session IDs and CDP endpoints.

kernel-agent-browser