The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install the agent-browser tool from the Vercel Labs GitHub repository. Vercel Labs is a trusted organization, and the download is intended for the skill's primary function of market research.
[REMOTE_CODE_EXECUTION]: The setup instructions include using npx to add external skills and install CLI tools (agent-browser). This executes code from a remote source, although it targets a trusted repository.
[COMMAND_EXECUTION]: The skill executes local shell scripts (tool-discovery.sh, chrome-profiles.sh) for environment discovery and uses CLI tools like agent-browser to interact with websites. These scripts are part of the vendor's provided marketing suite.
[DATA_EXFILTRATION]: The skill includes patterns for accessing persistent browser profiles (e.g., ~/.linkedin-profile) and saving authentication state files. These files contain sensitive session tokens. The skill documentation explicitly warns the user to add these files to .gitignore to prevent accidental exposure.
[PROMPT_INJECTION]: The skill's primary workflow involves using browser automation to scrape and analyze content from live websites, such as LinkedIn and competitor landing pages. This creates a surface for indirect prompt injection, where malicious instructions embedded in external web content could influence the agent's behavior.
Ingestion points: Web page content retrieved via agent-browser get text body and WebFetch.
Boundary markers: Not explicitly defined in the prompts.
Capability inventory: File system writes (saving reports), browser interaction (click, fill, screenshot), and shell script execution.
Sanitization: No specific sanitization or filtering of ingested HTML/text is documented.

paw-mkt-guerrilla