yc-apply
Pass
Audited by Gen Agent Trust Hub on Sep 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFECOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting untrusted data from external sources. These include external research findings generated by subagents from the web and coding-agent transcripts from local directories. The skill processes this data without explicit sanitization or boundary markers, creating a risk that instructions embedded in the external content could influence the agent's context or downstream actions.
- [CREDENTIALS_UNSAFE]: The skill facilitates the collection and processing of sensitive data, including product login credentials (username/password) required for the application. Additionally, it targets sensitive local paths such as
~/.claude/projects/to extract coding-agent transcripts, which may contain private project data or communication history. - [COMMAND_EXECUTION]: The skill automates several tasks through the execution of bundled shell scripts (
bootstrap-workspace.sh,buzzword-scan.sh,phase-status.sh,journal-append.sh). While these scripts appear to perform legitimate automation using standard system utilities, they represent a significant local execution capability that could be abused if the agent is compromised via prompt injection. - [DYNAMIC_EXECUTION]: The skill employs a dynamic multi-agent architecture, spawning parallel subagents to perform research slices. This dynamic execution model expands the attack surface by creating multiple autonomous contexts that operate on untrusted external data.
Audit Metadata