jev-use
Warn
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the 'Codex Computer Use' and 'cua' (Computer Use Agent) interfaces to programmatically interact with the host operating system's GUI. It performs actions such as searching for UI elements and executing inputs, which grants the agent significant control over the user's environment.
- [DYNAMIC_EXECUTION]: In
references/runtime.md, the skill performs dynamic loading of JavaScript modules usingawait importwith paths constructed at runtime (e.g.,pathToFileURL(repoDir + "/scripts/loop.mjs")). This pattern allows for the execution of code residing in the local repository based on computed paths. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text ingested from the system's Accessibility (AX) tree. Because the 'Jev' decision engine relies on this UI text to determine the next action, malicious content in a browser window, email, or chat application could act as an indirect injection, tricking the agent into executing unintended commands.
- [DATA_EXFILTRATION]: The skill extracts text-based content and metadata from the GUI to send it to the 'Jev' external model for decision-making. While the instructions state that screenshots are not sent, sensitive information visible in application titles, input fields, or document content is still captured and transmitted externally.
Audit Metadata