The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses ssh and sudo to perform remote and local administrative tasks. Notably, in references/INSTALLATION.md, it automatically appends public keys to the root user's authorized_keys file to enable automated management, which represents a significant modification of system access controls.
[REMOTE_CODE_EXECUTION]: Several files, including references/coolify.md and references/oci.md, instruct the agent to fetch and execute shell scripts directly from the internet via curl | bash. These scripts originate from various sources including cdn.coollabs.io and a placeholder GitHub repository (YOUR-ORG/vibeskills-demo).
[EXTERNAL_DOWNLOADS]: The skill downloads binaries and installers from various unverified external sources, such as GitHub releases (e.g., contabo/cntb) and S3 buckets (e.g., kasm-static-content.s3.amazonaws.com).
[CREDENTIALS_UNSAFE]: The skill is designed to manage and retrieve sensitive cloud provider tokens (e.g., HCLOUD_TOKEN, CLOUDFLARE_API_TOKEN, DIGITALOCEAN_ACCESS_TOKEN) and SSH private keys. While it utilizes a secrets-management script from a sibling skill, it creates a centralized target for credential exposure.
[PROMPT_INJECTION]: The skill has a significant surface for indirect prompt injection as it ingests untrusted user parameters and interpolates them into shell commands across multiple provisioning scripts.
Ingestion points: User-provided parameters for server names, IP addresses, and application configurations (e.g., references/contabo.md, references/coolify.md).
Boundary markers: Absent; variables are directly interpolated into bash commands.
Capability inventory: Full shell access (bash), remote execution (ssh), and cloud administration (oci, hcloud, doctl, cntb, linode-cli) available in references/*.md and scripts/*.sh.
Sanitization: Absent; the skill relies on the user providing valid inputs without explicit escaping or validation of shell characters.

devops