The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructs the agent to use the mcp__node_repl__js tool to dynamically import the @oai/sky runtime using await import("@oai/sky"). This pattern allows for the execution of arbitrary code within the Node.js environment.- [COMMAND_EXECUTION]: The instructions require the agent to send Javascript code as single-line strings to the mcp__node_repl__js tool to interact with the computer use runtime.- [PRIVILEGE_ESCALATION]: The skill is designed to grant the agent high-privilege capabilities, including the ability to operate desktop application UIs, use browsers like Safari or Chrome, perform click and type actions, and capture screenshots of the user's screen.- [INDIRECT_PROMPT_INJECTION]: The agent is directed to read a local file (~/.codex/plugins/cache/openai-bundled/computer-use/<version>/skills/computer-use/SKILL.md) and treat its content as authoritative instructions. This creates a vulnerability surface where the agent's behavior is governed by external, potentially mutable local data.
Ingestion points: Local path ~/.codex/plugins/cache/openai-bundled/computer-use/<version>/skills/computer-use/SKILL.md identified in SKILL.md.
Boundary markers: Absent; the agent is explicitly told the external skill is 'authoritative'.
Capability inventory: Includes arbitrary Javascript execution via nodeRepl.write, UI automation, and desktop screenshotting.
Sanitization: Absent; the agent is instructed to read the file content directly before performing computer-use work.

codex-computer-use