continue-on-phone
Warn
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads the
acpantennaandqrcode-terminalpackages from the NPM registry at runtime usingnpx -y. - [REMOTE_CODE_EXECUTION]: Packages downloaded via
npxare executed immediately to facilitate the remote connection and generate QR codes. - [COMMAND_EXECUTION]: The skill initiates background shell processes, creates temporary directories, and uses redirection to pipe inputs and outputs between the local agent session and the remote relay.
- [INDIRECT_PROMPT_INJECTION]:
- Ingestion points: Prompts and commands are received from a remote browser or mobile device via an external relay service (
latch.im). - Boundary markers: Uses a four-word pairing protocol that requires the human user to manually verify and confirm the connection match between the terminal and browser.
- Capability inventory: The remote browser acts as a client that can prompt the agent and trigger its local tools, including filesystem access and command execution in the project directory.
- Sanitization: Employs an ACP filtering proxy that limits certain request types (like 'always allow' permissions) and filters specific file/terminal requests.
Audit Metadata