The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script scripts/unifi-firewall-score to calculate security scores from JSON input. It also uses shell commands (mkdir, echo, python3) to manage an audit history file.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from the UniFi controller, including firewall rule names, descriptions, and device names. This data is incorporated into findings and the final audit report, creating a potential surface for indirect prompt injection if the agent interprets malicious instructions embedded in these strings.
Ingestion points: Network and firewall metadata retrieved via unifi_list_firewall_policies, unifi_list_networks, and unifi_list_devices in SKILL.md.
Boundary markers: None explicitly used when interpolating controller-provided strings into findings or the final report.
Capability inventory: Subprocess execution for scoring (scripts/unifi-firewall-score) and shell commands for state management.
Sanitization: No filtering or sanitization of rule names or descriptions is observed before they are processed by the agent.
[PERSISTENCE]: The skill maintains a local audit log at ~/.local/state/unifi-mcp/skills/audit-history.json. It appends new audit results and trims the file to the last 50 entries to track security trends over time.
[DYNAMIC_EXECUTION]: The skill uses shell-based string interpolation to generate and execute Python one-liners for history management. While the data being interpolated (scores and timestamps) is currently constrained, this pattern of dynamic script generation can be risky if external inputs are not strictly validated.

firewall-auditor