The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill requires the installation of parallel-cli via curl -fsSL https://parallel.ai/install.sh | bash. This pattern executes a remote shell script from a third-party domain (parallel.ai) that is not included in the trusted vendor list, posing a significant risk of arbitrary code execution.\n- [INDIRECT_PROMPT_INJECTION]: The skill synthesizes data from external research papers and web results, creating a surface for indirect prompt injection where malicious instructions in academic content could influence agent behavior.\n
Ingestion points: Data retrieved from PubMed, arXiv, and general web searches via parallel-cli search, stored in the sources/ directory.\n
Boundary markers: Absent. The skill does not use delimiters or explicit instructions to isolate untrusted research data within the prompt context during the synthesis phase.\n
Capability inventory: Access to the Bash tool, file system write access, and execution of system tools like pandoc through scripts/generate_pdf.py.\n
Sanitization: Absent. There is no logic provided to sanitize, filter, or escape academic text before it is interpolated into prompts or synthesized into the final review document.\n- [COMMAND_EXECUTION]: The scripts scripts/generate_pdf.py and scripts/generate_schematic.py utilize subprocess.run to execute external binary tools such as pandoc and xelatex, as well as child Python scripts, which could be exploited if arguments are not properly handled.

literature-review