learn
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONPERSISTENCE
Full Analysis
- [COMMAND_EXECUTION]: The skill provides explicit instructions for the agent to execute shell commands, including repository status checks (
git status), cloning repositories (git clone), and a full software build pipeline usingpnpm installandpnpm turbo build. - [DYNAMIC_EXECUTION]: The skill features a mechanism to fetch and "follow" remote instruction files (
SKILL.md) fromratstack.shat runtime, which allows the agent to adopt new behaviors based on external content. It also directs the agent to execute a locally built JavaScript artifact as a persistent MCP tool. - [EXTERNAL_DOWNLOADS]: The skill fetches learning decks and configuration data from
ratstack.shand clones source code fromgithub.com/joelhooks/rat-stack.git. These resources are owned by the skill author. - [INDIRECT_PROMPT_INJECTION]: The ingestion of remote learning decks and additional skill files from
ratstack.shcreates a surface for potential instruction injection where external content could influence agent actions. - Ingestion points: Remote JSON decks and Markdown instruction files fetched via network requests to
ratstack.sh. - Boundary markers: There are no instructions provided to treat the fetched content as data only or to ignore embedded commands within the dynamic skills.
- Capability inventory: The agent has access to shell execution, package managers, and network utilities.
- Sanitization: The skill does not define validation or sanitization steps for the remotely fetched instructions before the agent is told to "follow" them.
- [PERSISTENCE]: The instructions direct the agent to configure its stdio MCP client to include the locally built tool, which ensures the tool and its capabilities persist across future agent sessions.
Audit Metadata