The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from vCenter events and ESXi host logs. It implements robust security measures to prevent this data from influencing the agent's behavior.
Ingestion points: The skill reads vCenter alarms, events, and host logs (hostd, vmkernel, vpxa) via the list_vcenter_alarms tool and the background scanning daemon.
Boundary markers: External content is wrapped in explicit [VSPHERE_EVENT] and [VSPHERE_HOST_LOG] delimiters.
Capability inventory: The skill has significant capabilities, including 43 write tools such as vm_delete, vm_guest_exec, and cluster_delete.
Sanitization: vSphere-sourced content is truncated, stripped of control characters, and sanitized before being processed by the agent.
[COMMAND_EXECUTION]: The skill provides tools for executing arbitrary commands inside target guest VMs (vm_guest_exec, vm_guest_exec_output). This is a documented core functionality for managing VMware guests and is mitigated by a separate authorization boundary (guest OS credentials) and explicit documentation advising the use of least-privilege accounts.
[EXTERNAL_DOWNLOADS]: The skill's installer and dependencies are fetched from official package registries (PyPI) and the vendor's own GitHub repositories. This behavior is documented and follows standard installation practices.
[DATA_EXFILTRATION]: The skill supports user-configured webhooks for Slack and Discord. The documentation notes that these are disabled by default, only send to user-configured endpoints, and include sanitized operational data without credentials.
[CREDENTIALS_UNSAFE]: The skill manages credentials using local .env files. It employs a basic Base64-based obfuscation layer for passwords stored at rest to prevent accidental exposure via local file searches. This is presented as a local security feature and follows standard secret management practices for this platform.

vmware-aiops