The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses execFileSync to drive ffmpeg and ffprobe for video transcoding and metadata verification. These calls use fixed argument arrays and are limited to processing local files generated by the recorder, preventing arbitrary command injection.
[REMOTE_CODE_EXECUTION]: The setup.sh script installs pinned dependencies (playwright and yaml) and Chromium via official registries (npm and npx playwright). These are standard development operations from trusted sources.
[DATA_EXFILTRATION]: While the skill records browser content, it includes a fail-closed privacy masking system. If a requested mask selector is not found, the recording aborts immediately to prevent accidental exposure of sensitive data. It also warns users to confirm origins for authenticated recordings.
[PRIVILEGE_ESCALATION]: No use of sudo, chmod 777, or other privilege escalation techniques were detected in the setup or runtime scripts.
[DYNAMIC_EXECUTION]: The skill injects a local script (overlay.js) into the browser context to render cursors and cards. This script is bundled with the skill and does not load external code at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted app URLs and scenario files. However, it operates in a standalone browser context separate from the agent's environment, and the YAML validator prevents malformed or malicious structures from reaching the execution engine.

walkthrough-record