full-feature-flow
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from repository files and instructions, creating an attack surface for indirect prompt injection.
- Ingestion points: The orchestrator in
SKILL.md(Step 1) gathers repository context and files, which are subsequently read and processed by thePlannerandCoderroles. - Boundary markers: There are no specific instructions or delimiters (such as XML tags or triple-backticks with explicit ignore-directives) mandated to separate untrusted file content from the agent's instructions.
- Capability inventory: The workflow grants significant capabilities to the agents, including the ability for the
Coderto edit the working tree, start web servers, and execute arbitrary CLI commands for testing and building. Additionally, Reviewer roles are authorized to execute "short scripts" and make network requests to running applications. - Sanitization: The skill does not define any sanitization or validation logic for the content ingested from the repository before it is presented to the LLM-based agents.
- [COMMAND_EXECUTION]: Several roles in the workflow are explicitly designed to execute commands on the host system.
- Evidence: The
Coderrole inreferences/coder.mdis responsible for starting web apps, running project test suites, and executing CLI commands. The Reviewer roles defined inreferences/code-reviewer.mdandreferences/critical-reviewer.mdare permitted to run "short scripts" or "targeted checks" to verify findings.
Audit Metadata