computer-use
Warn
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: MEDIUMPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The instructions explicitly guide the agent to escalate its permissions when encountering sandbox-related access denials.
- Evidence: The text in
SKILL.mdinstructs the agent that if a command fails withruntime_access_denied, it should "re-run it with escalated permissions". - [INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface for malicious instructions embedded in external application windows to influence agent behavior.
- Ingestion points: GUI accessibility trees, screenshots, and window contents from native apps and external web browsers (Chrome, Edge, Safari) through the
orca computertool referenced inSKILL.md. - Boundary markers: Absent. There are no instructions to use delimiters or ignore instructions found within the processed GUI data.
- Capability inventory: Execution of system commands and full GUI manipulation capabilities (clicks, typing, menus).
- Sanitization: Absent. Data retrieved from the system interface is not validated or sanitized.
- [DYNAMIC_EXECUTION]: The binary used for system interaction is resolved at runtime using environment variables.
- Evidence: The logic in
SKILL.mddetermines theORCAexecutable based onORCA_CLI_COMMANDandORCA_DEV_REPO_ROOT. - [COMMAND_EXECUTION]: The skill facilitates the execution of shell commands to discover and interact with system tools.
Audit Metadata