The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from web server and WAF logs which may contain malicious payloads designed to influence agent behavior. It proactively mitigates this risk through explicit instructions.
Ingestion points: Log retrieval via docker logs, kubectl logs, and journalctl (found in SKILL.md).
Boundary markers: The skill includes a 'Data, not instructions' section that explicitly warns the agent about attacker-controlled content in logs.
Capability inventory: Administrative capabilities including docker exec, kubectl exec, and sudo for managing the firewall and system services.
Sanitization: Employs prompt-level sanitization by instructing the agent to treat quoted request data inside any log line strictly as data, never as instructions.
[COMMAND_EXECUTION]: The skill provides commands to execute administrative tasks within containerized and host environments.
Evidence: Instructions for docker exec, kubectl exec, and direct execution of the bwcli tool to list or lift IP bans.
Context: These commands are limited to the vendor's own CLI tools and standard diagnostic utilities necessary for the skill's primary troubleshooting purpose.
[PRIVILEGE_ESCALATION]: The skill mentions the use of elevated privileges for system-level troubleshooting on Linux installations.
Evidence: Use of sudo for bwcli bans and systemctl restart bunkerweb-scheduler (found in SKILL.md).
Context: This is standard practice for the administration of system services and firewall software and aligns with the expected behavior for this vendor's tools.

bunkerweb-troubleshooting