The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates browsing and interacting with external websites. Because the agent ingests content from these pages into its context, it is exposed to potential indirect prompt injection attacks where malicious web content could attempt to override agent instructions.
Evidence: Use of page.goto and interactive browser driving in scripts/record-flow.mjs via the agent-browser dependency.- [PRIVILEGE_ESCALATION]: Documentation in SKILL.md suggests that an agent may need to remount system resources (/dev/shm) if Chromium crashes in restricted environments. Such operations typically require administrative privileges, and prompting an agent to perform system-level modifications is a security concern.
Evidence: Troubleshooting guidance in SKILL.md regarding checking and remounting /dev/shm.- [COMMAND_EXECUTION]: The skill relies on several shell commands for installation and operation, including Node.js package management and browser driver setup.
Evidence: Instructions in SKILL.md and logic in scripts/setup.sh.- [DATA_EXFILTRATION]: By recording the browser viewport, the skill captures screenshots and video of all on-screen activity. If the user or agent navigates to sensitive sites during a recording, private data or credentials could be captured and stored on the host filesystem.
Evidence: Implementation of Playwright's screenshot and recordVideo features in scripts/record-flow.mjs.

record-web-journey