unvibe
Pass
Audited by Gen Agent Trust Hub on Jun 21, 2026
Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes standard development commands including
git diff,git show, and various test runners such aspytest,go test, andnpx vitest. These operations are conducted within the local workspace to analyze project changes and verify user solutions, which is consistent with the skill's stated purpose. - [REMOTE_CODE_EXECUTION]: The agent generates local exercise files and shell scripts (
run.sh) based on included templates and subsequently executes them. This dynamic execution is part of the core functionality to provide a runnable testing environment. All generated content is derived from the skill's own assets and executed locally. - [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting untrusted code from the local filesystem and git history (e.g.,
git diffoutput). While this could theoretically be used to influence agent behavior, the risk is minimal due to the diagnostic nature of the skill. Evidence chain: (1) Ingestion: Git logs and source files; (2) Boundaries: None; (3) Capabilities: Shell execution, file writes, and WebFetch; (4) Sanitization: None. - [DATA_EXFILTRATION]: The skill uses
WebSearchandWebFetchto retrieve technical documentation from well-known resources (like MDN) to verify facts during exercise creation. This network activity is limited to information retrieval and does not involve the exfiltration of sensitive data.
Audit Metadata