The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill's installation script attempts to gain administrative privileges and bypass system security settings.
Evidence: In scripts/install.sh, the run_with_sudo_if_needed function is used to execute rm, mv, and installer commands as root.
Evidence: In scripts/install.sh, the script executes xattr -dr com.apple.quarantine to remove macOS Gatekeeper protection from the downloaded application.
[EXTERNAL_DOWNLOADS]: The skill downloads binary artifacts from an external vendor domain during setup.
Evidence: In scripts/install.sh, curl is used to fetch .dmg files from https://cdn.ego.app/setup/macos/arm64/egolite-fHoqgZ74bOEM.dmg and https://cdn.ego.app/setup/macos/x64/egolite-fHoqgZ74bOEM.dmg.
[REMOTE_CODE_EXECUTION]: The skill downloads and executes an external application on the host system.
Evidence: scripts/install.sh downloads a disk image (DMG), mounts it using hdiutil, and installs the contained application to the /Applications directory.
[COMMAND_EXECUTION]: The skill utilizes shell commands to manage its environment and execute browser automation logic.
Evidence: SKILL.md provides examples of using ego-browser nodejs with heredocs to execute arbitrary Node.js code.
Evidence: references/install.md instructs the user to run the installation script using sh skills/ego-browser/scripts/install.sh.
[INDIRECT_PROMPT_INJECTION]: The skill interacts with and extracts data from arbitrary websites, creating a path for malicious content to influence agent behavior.
Ingestion points: page.snapshot() and page.evaluate() in SKILL.md extract content from untrusted web pages into the agent context.
Boundary markers: The skill instructions do not specify the use of delimiters or 'ignore' instructions for data processed from the browser.
Capability inventory: The skill can perform file writes (download.saveAs), network requests (page.fetch), and execute shell commands through the ego-browser CLI.
Sanitization: There is no evidence of input validation or output escaping for data extracted from web pages before it is processed by the agent.
[DYNAMIC_EXECUTION]: The skill enables the agent to generate and execute code at runtime.
Evidence: SKILL.md documents the page.evaluate(fnOrString) method, which executes arbitrary JavaScript within the browser context.
Evidence: The core functionality relies on passing JavaScript strings to the ego-browser nodejs command for execution.

ego-browser