The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The SKILL.md file includes instructions to install dependencies via curl -fsSL https://parallel.ai/install.sh | bash. This delivery method allows for arbitrary code execution from a remote server without verification and is classified as a critical risk.
[COMMAND_EXECUTION]: The script scripts/generate_schematic.py uses subprocess.run() to execute other local scripts. This represents a command execution surface that could be exploited to run arbitrary code on the local machine.
[DATA_EXFILTRATION]: The skill transmits research queries to api.parallel.ai and openrouter.ai. These domains are not part of the established whitelist of trusted sources, posing a risk of sensitive data exposure through non-whitelisted network operations.
[PROMPT_INJECTION]: The skill ingests untrusted search data from external APIs, creating a vulnerability to indirect prompt injection.
Ingestion points: Web and academic search snippets from Parallel AI and Perplexity.
Boundary markers: No explicit markers are used to separate ingested data from agent instructions.
Capability inventory: Access to shell execution (subprocess.run), file system (Read/Write/Edit), and network (requests).
Sanitization: No sanitization or filtering of external content is performed before processing.

research-lookup