The Agent Skills Directory

PROMPT_INJECTION (HIGH): The skill is susceptible to indirect prompt injection as it processes untrusted markdown content that is subsequently used in browser automation tasks.
Ingestion points: The scripts scripts/format-for-x.sh and scripts/publish-article.sh ingest user-provided content files without sanitization.
Boundary markers: No delimiters or instructions are present to prevent the agent from obeying command sequences embedded in the article text.
Capability inventory: The skill utilizes agent-browser which permits arbitrary navigation, button interaction, file uploads, and JavaScript execution (evaluate).
Sanitization: The formatting script only removes markdown syntax and does not filter or escape natural language instructions that could trigger agent actions.
COMMAND_EXECUTION (MEDIUM): The skill invokes external CLI tools (agent-browser and pbcopy) to perform complex browser interactions.
Evidence: scripts/publish-article.sh passes user-controlled file paths to these tools and uses dynamic JavaScript execution via the evaluate flag to interact with the DOM.
EXTERNAL_DOWNLOADS (LOW): The skill relies on an unverified third-party tool agent-browser.
Evidence: skill.json lists agent-browser as a required dependency. While common for automation agents, the source repository is not in the trusted list and requires external installation.

x-articles