The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/perplexity_search.sh performs unsafe dynamic code generation by interpolating shell variables into a Python command.
Evidence: The script assembles a Python payload using BODY=$(python3 -c "... 'content': '''$QUERY''' ...").
Risk: Because $QUERY is interpolated directly into Python triple-quotes without escaping, a malicious input containing triple-quotes (''' ) can break out of the string literal and execute arbitrary Python commands on the user's system.
[PROMPT_INJECTION]: The skill exhibits misleading metadata and lacks safeguards against indirect prompt injection.
Evidence (Metadata): The SKILL.md documentation provides a guide for multiple models like sonar-reasoning-pro and sonar-research, but the scripts/perplexity_search.sh script contains a hardcoded check that terminates execution if any model other than the base sonar is used.
Evidence (Indirect Injection Surface):
Ingestion points: The $QUERY and $SYSTEM_PROMPT variables in scripts/perplexity_search.sh accept untrusted input.
Boundary markers: No delimiters or instruction-ignore wrappers are used when passing these inputs to the API or the local Python execution block.
Capability inventory: The skill possesses network access via curl and local script execution via python3.
Sanitization: No sanitization, escaping, or validation is performed on inputs before they are used in command-line arguments and API payloads.

perplexity-4