The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's setup process and documentation recommend installing the Bright Data CLI using a shell script fetched from a remote URL (https://cli.brightdata.com/install.sh) and piped directly to the system shell. While this originates from a well-known service provider, this execution pattern is a common vector for supplying external code to the environment.
[COMMAND_EXECUTION]: Several components, including scripts/bd_search.py and scripts/research_engine.py, use subprocess.run to execute command-line tools. The search wrapper dynamically builds shell commands based on user-supplied queries and URLs to interface with the brightdata CLI. While it avoids shell=True in key areas, it remains an active capability for system-level interaction.
[PROMPT_INJECTION]: The skill is designed to ingest large amounts of untrusted data from the open web, including search results, scraped articles, and social media posts (e.g., via the Reddit pipeline). This creates a vulnerability to indirect prompt injection, where an attacker could place malicious instructions on a web page that, when processed by the agent during the synthesis or critique phases, could override its behavior or bias the generated research report.

deep-research