ubicloud
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill relies on several local binaries located in
~/.claude/skills/gstack/bin/to manage session state, telemetry, and VM lifecycle (e.g.,gstack-skill-start,gstack-ubi,gstack-skill-end). These commands are executed with various arguments and environmental flags. - [REMOTE_CODE_EXECUTION]: By design, this skill's primary function is to execute code on remote infrastructure (Ubicloud VMs). It automates the creation of VMs, streams local file content to them, and executes bootstrap and test scripts.
- [INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection via the
.gstack/ubicloud.jsonfile in the user's repository. This file defines thesetupscript andcommandto be executed on the VM. - Ingestion points:
.gstack/ubicloud.jsonfile (SKILL.md Step 3). - Boundary markers: The instructions explicitly tell the agent to "Run verify only on a tree you have reviewed."
- Capability inventory: The skill can execute arbitrary commands on a remote VM, pass environment variables, and pull files back from the remote VM to the local host.
- Sanitization: The runner refuses absolute or parent directory (
..) paths for script locations and file pull destinations to prevent directory traversal. - [DYNAMIC_EXECUTION]: The skill implements a mechanism for dynamic instruction injection. The preamble instructs the agent to look for and follow
GSTACK_INSTRUCTION_BEGINblocks within the output of thegstack-skill-starttool. These blocks can contain one-time onboarding or consent directives that the agent is told to follow before proceeding.
Audit Metadata