The Agent Skills Directory

[SAFE]: The skill demonstrates high security awareness through multiple defensive layers and careful handling of external data.
[COMMAND_EXECUTION]: The scripts/campaign_manager.py script generates shell command strings based on user-provided templates. This is a common pattern for orchestration but is handled safely:
A _validate_command_template function checks for and rejects dangerous shell operators such as pipes, semicolons, and backgrounding operators.
Configuration paths are validated against a safe-character allowlist (_SAFE_PATH_PATTERN).
Shell arguments are escaped using shlex.quote() to prevent injection.
The skill's allowed-tools configuration in SKILL.md explicitly excludes Bash, ensuring the agent cannot execute the generated commands itself.
[INDIRECT_PROMPT_INJECTION]: The scripts/result_aggregator.py script processes external simulation result files, which represents a potential indirect prompt injection surface. The script implements comprehensive mitigations:
Ingestion points: Processes JSON result files (e.g., result_*.json) produced by external simulations.
Sanitization: The _sanitize_value() function truncates strings to 500 characters and strips non-printable control characters, preventing large or malicious payloads from entering the agent context.
Validation: The script enforces strict limits: a 10MB maximum file size, a maximum JSON nesting depth of 10, and a regex-validated pattern for metric names.
Capability inventory: The skill is limited to local file system operations (Read/Write) and cannot perform network operations or direct command execution.

simulation-orchestrator