The Agent Skills Directory

[DATA_EXFILTRATION]: The skill transmits desktop screen metadata, including element names, roles, and partial values, to an external API (api.typesafe.ai) for processing. The documentation explicitly notes that this can include the contents of private documents or filled forms.
Evidence: SKILL.md states: "Every turn posts a description of the screen to api.typesafe.ai: each element's role, its accessible name or description, up to sixty characters of the value it holds..."
[INDIRECT_PROMPT_INJECTION]: The skill processes natural language goals to perform desktop automation, creating a surface for potential injection attacks where malicious content on the screen could influence the agent's actions.
Ingestion points: Natural language intents are provided via command-line arguments in scripts/jev/run.mjs and scripts/jev/act.mjs. The skill also reads external data from the desktop screen.
Boundary markers: The skill does not define explicit boundary markers or delimiters for the ingested goals or screen content.
Capability inventory: The skill can perform various desktop interactions including CLICK, TYPE_TEXT, and SCROLL via scripts/jev/desktop.mjs which interfaces with the agent-desktop tool.
Sanitization: The skill employs confidence-based gates (0.70 for standard actions, 0.90 for destructive ones) and provides a --no-values flag to restrict data exposure, which serves as a partial mitigation.

jev-desktop