oma-hwp
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill relies on
bunx kordoc@latestto download and execute the conversion utility from the NPM registry at runtime. While NPM is a well-known service, executing unpinned packages (@latest) from third-party authors introduces risks of supply chain attacks or breaking changes. - [REMOTE_CODE_EXECUTION]: In
resources/execution-protocol.md, the skill provides recovery instructions involvingcurl -fsSL https://bun.sh/install | bash. This is a classic remote script execution pattern; although it targets the official well-known Bun installation service, it bypasses local security controls. - [COMMAND_EXECUTION]: The skill executes shell commands via
bunandbunxto run the converter and a post-processing script (resources/flatten-tables.ts). These commands use variables ({input_path},{output_path}) which, if not properly handled by the agent, could lead to command injection. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted Korean document files (.hwp, .hwpx) and convert them to Markdown for consumption by the agent.
- Ingestion points: HWP-family files provided via
input_path(specified inSKILL.md). - Boundary markers: The instructions do not specify using delimiters or "ignore" warnings when the agent reads the converted Markdown output.
- Capability inventory: The skill has filesystem write access and the ability to execute subprocesses (
bunx kordoc,bun flatten-tables.ts). - Sanitization: The
kordoctool performs link sanitization (perresources/troubleshooting.md), and theflatten-tables.tsscript strips Hancom-specific Private Use Area (PUA) characters to prevent rendering issues.
Audit Metadata