The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the $TASK environment variable directly within shell hooks, which is a potential command injection vector.
Evidence: The pre hook in SKILL.md executes echo "📊 Performance Benchmarker analyzing: $TASK". If the task name provided to the agent contains shell metacharacters (e.g., $(dangerous_command)), they could be executed by the shell interpreter.
[DATA_EXFILTRATION]: The skill collects and persists sensitive system-level resource metrics.
Evidence: The ResourceUsageMonitor class implements methods to collect high-fidelity telemetry including totalUsage, loadAverage, processRSS, networkIO (bytes in/out), and diskIO details via a SystemMonitor utility.
Evidence: The collected system state is stored for up to 7 days using mcpTools.memory_usage and persisted locally in historicalData.store.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted configuration data to drive benchmarking and system optimization logic without sufficient sanitization.
Ingestion points: External data enters the context through the $TASK environment variable and the scenarios parameter passed to the runComprehensiveBenchmarks function in SKILL.md.
Boundary markers: The skill does not implement delimiters or instructions to the agent to disregard instructions that might be embedded within scenario descriptions or task names.
Capability inventory: The skill possesses extensive system monitoring capabilities and the ability to modify system/protocol parameters through the AdaptiveOptimizer class.
Sanitization: There is no evidence of validation or filtering for the scenarios data before it is processed to initialize benchmark environments and execute tests.

agent-performance-benchmarker