The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The installation instructions guide the user to execute uvx with a remote GitHub repository (mpeirone/zabbix-mcp-server). This pattern downloads and runs unverified external code on the user's system at runtime.
[EXTERNAL_DOWNLOADS]: The skill relies on fetching code from an external, non-authoritative repository during installation and execution, creating a dependency on third-party code that has not been audited.
[COMMAND_EXECUTION]: The tool zabbix_api provides the agent with the ability to execute any method on the Zabbix API. Without user-configured restrictions such as READ_ONLY, ZABBIX_API_WHITELIST, or ZABBIX_API_BLACKLIST, the agent can perform destructive actions like deleting hosts, items, or triggers.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through data retrieved from the Zabbix API. \n
Ingestion points: Data fields like host names, problem descriptions, and trigger expressions retrieved via the zabbix_api tool. \n
Boundary markers: No specific delimiters or instructions are provided to the agent to treat API responses as untrusted data. \n
Capability inventory: The zabbix_api tool allows for creating and modifying Zabbix resources, providing an attacker with a way to influence the monitoring environment if they can inject instructions into the monitoring data. \n
Sanitization: There is no evidence of sanitization or filtering of the Zabbix API output before it is processed by the AI agent.

zabbix-mcp-server