computer-use
Fail
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: HIGHPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill explicitly directs the agent to "re-run it with escalated permissions" if a
runtime_access_deniederror occurs. This instructs the agent to bypass security or sandbox constraints by acquiring administrative or root-level privileges. - [DYNAMIC_EXECUTION]: The skill acts as a "discovery stub," meaning its core instructions are not present in the skill file. Instead, it uses the command
ORCA skills get computer-useto fetch and load a "version-matched guide" at runtime. This dynamic loading of instruction content bypasses static analysis and allows the behavior of the skill to change without updating the source file. - [COMMAND_EXECUTION]: The skill provides logic for the agent to identify and execute various local binaries (
orca,orca-ide,orca-dev) based on the environment. This facilitates direct interaction with the host operating system and local application GUI. - [INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface by instructing the agent to follow a guide fetched from a command output.
- Ingestion points: Content returned from
ORCA skills get computer-useinSKILL.md. - Boundary markers: Absent; there are no instructions to treat the fetched guide as data or to ignore embedded instructions.
- Capability inventory: Native GUI control (accessibility tree, clicks, typing, menus), screenshotting, and browser interaction.
- Sanitization: Absent; the agent is expected to directly "load" and follow the guide returned by the executable.
Recommendations
- AI detected serious security threats
Audit Metadata