qmd
Fail
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSPERSISTENCEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONTIME_DELAYED_CONDITIONAL
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill setup (
setup/README.md) and diagnostic script (scripts/doctor.sh) suggest installing the Bun runtime via a shell script from its official domain (https://bun.sh/install). It also installs theqmdtool directly from thetobi/qmdGitHub repository using the Bun package manager. - [PERSISTENCE]: The skill includes a macOS LaunchAgent template (
templates/launchd.plist) and instructions for its installation to perform daily background embedding refreshes. It also recommends configuringPostToolUsehooks in the agent'ssettings.jsonto trigger real-time index updates after file modifications. - [DYNAMIC_EXECUTION]: The diagnostic script
scripts/doctor.shuses theevalcommand to perform status checks on system dependencies. This allows the script to dynamically verify the presence of tools like Bun, SQLite, and theqmdbinary in the user's environment. - [TIME_DELAYED_CONDITIONAL]: The automation strategy relies on a
launchdjob scheduled to run daily at midnight to regenerate embeddings. This time-gated execution manages resource-heavy indexing tasks outside of active working hours. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to search and retrieve content from a local collection of markdown files, which represents a potential attack surface.
- Ingestion points: Markdown content is ingested through search and query commands (
qmd search,qmd vsearch,qmd query) from files in the user's project directory as specified inSKILL.mdandsetup/README.md. - Boundary markers: There are no explicit markers or prompt instructions to isolate search results from the agent's primary instructions or to treat the retrieved content as untrusted.
- Capability inventory: The agent has access to
Bash,Read,Write, andEdittools, which could be misused if search results contain malicious instructions that the agent executes. - Sanitization: The skill does not perform sanitization or validation of the indexed markdown content before presenting it to the agent.
- [COMMAND_EXECUTION]: The skill frequently executes external commands, including the
qmdCLI for search and indexing, and system utilities likelaunchctl,brew, andfswatchfor environment management and automation.
Recommendations
- HIGH: Downloads and executes remote code from: https://bun.sh/install - DO NOT USE without thorough review
Audit Metadata