The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from Proxmox nodes and guest environments to drive high-privilege administrative tasks, creating a surface for indirect manipulation.
Ingestion points: Reads PVE API documentation and live schemas from node-served endpoints (https://<node>:8006/pve-docs/api-viewer/apidoc.js) and captures guest-level output over SSH during provisioning (bake step in references/golden-image.md).
Boundary markers: The instructions lack explicit delimiters or warnings for the agent to ignore potentially malicious content within the ingested node or guest output.
Capability inventory: The skill utilizes ssh, curl, and package management tools (apt) which could be exploited if malicious data influences the agent's logic.
Sanitization: No sanitization or validation logic is specified for data retrieved from external nodes before its use in command interpolation.
[PRIVILEGE_ESCALATION]: The skill provides instructions for performing host-level modifications and executing commands with elevated privileges.
Evidence: Direct modification of system repository files in /etc/apt/sources.list.d/ using sed -i to change release suites as described in references/upgrade-8-to-9.md.
Evidence: Execution of host-level major upgrades using apt dist-upgrade and the pve8to9 utility.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill automates the retrieval and installation of software packages and OS images from remote sources.
Evidence: Fetches updates and packages from the official Proxmox repository (http://download.proxmox.com) during the 8-to-9 upgrade sequence.
Evidence: Downloads Ubuntu cloud images from Canonical sources via the PVE download-url API for golden image templates.
Evidence: Recommends the use of curl -k (or --insecure) when interacting with PVE nodes, which is a common practice for handling self-signed certificates in hypervisor environments.

proxmox