dbs-computer-use
Fail
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill provides explicit instructions to download and execute shell scripts directly from external domains cua.ai and opencode.ai. These scripts are executed via shell pipes (curl | bash, curl | sh) or PowerShell's iex command, which bypasses file-based security scanning.
- Evidence in references/setup.md: /bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)"
- Evidence in references/setup.md: irm https://cua.ai/driver/install.ps1 | iex
- Evidence in references/setup.md: curl -fsSL https://opencode.ai/browse/install | bash
- [EXTERNAL_DOWNLOADS]: The skill facilitates the installation of third-party software from non-standard sources. It directs the agent to install agent-desktop via the NPM registry and multiple specialized drivers from external sites.
- Evidence in references/setup.md: npm install -g agent-desktop
- [INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for indirect prompt injection because it reads and interacts with untrusted content from the web and desktop applications.
- Ingestion points: references/cua-driver.md (calls get_window_state and get_browser_state to read UI elements and web content), references/agent-desktop.md (uses snapshot to read window elements).
- Boundary markers: None identified. There are no instructions or delimiters used to ensure the agent ignores malicious prompts embedded in the processed application or web data.
- Capability inventory: The skill has high-privilege capabilities including executing shell commands, automating browsers, and simulating mouse/keyboard interactions.
- Sanitization: There is no mention of filtering or sanitizing the data read from external interfaces before it is processed by the agent.
- [DYNAMIC_EXECUTION]: The skill documentation acknowledges that one of its recommended components (OpenCode Browse) utilizes eval to execute arbitrary code within the browser context, which increases the risk if the component is compromised or receives malicious input.
Recommendations
- AI detected serious security threats
Audit Metadata