The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill explicitly directs the agent to bypass sandbox restrictions by re-running commands with "escalated permissions" if it encounters a runtime_access_denied error. This encourages the agent to seek higher system privileges (such as sudo or administrator access) to overcome security boundaries.
[DYNAMIC_EXECUTION]: The instructions mandate that the agent must load a "version-matched guide" at runtime using the command ORCA skills get orca-cli. This pattern results in the execution of instructions that are not statically defined in the skill, making the agent's behavior unpredictable and difficult to audit.
[INDIRECT_PROMPT_INJECTION]: By processing the output of the ORCA skills get command as a guide for further actions, the skill creates an injection surface. If the external CLI tool or the data it retrieves is compromised, it can provide malicious instructions that the agent will follow as if they were part of its core logic.
[COMMAND_EXECUTION]: The skill determines which system binary to execute (orca, orca-ide, or orca-dev) based on the presence of environment variables like ORCA_CLI_COMMAND and ORCA_DEV_REPO_ROOT, which could be manipulated to redirect execution to unauthorized or malicious binaries.

orca-cli