The Agent Skills Directory

[PROMPT_INJECTION]: The skill enables the agent to ingest data from external sources, including terminal output via orca terminal read and browser accessibility trees via orca snapshot. This creates an indirect prompt injection surface where untrusted content could potentially manipulate the agent's instructions. \n
Ingestion points: orca terminal read, orca snapshot (SKILL.md)\n
Boundary markers: The skill mentions a 'Boundary with orchestration' for agent-to-agent communication but lacks explicit delimiters or instructions to ignore embedded commands in raw data.\n
Capability inventory: orca terminal send, orca terminal create, orca eval, orca exec (SKILL.md)\n
Sanitization: No evidence of input validation or filtering for data processed from the terminal or browser.\n- [COMMAND_EXECUTION]: The skill facilitates shell command execution through orca terminal send and allows the creation of new terminals running arbitrary commands via orca terminal create --command.\n- [REMOTE_CODE_EXECUTION]: The skill documents the orca eval command, which permits the execution of arbitrary JavaScript within the context of the browser, potentially allowing for script injection or cross-site scripting (XSS) scenarios if the input is derived from untrusted sources.\n- [DATA_EXFILTRATION]: The skill exposes commands that can retrieve sensitive data, including orca cookie get for browser session cookies, orca screenshot for visual data, and orca terminal read for console history.

orca-cli