ace-step
Pass
Audited by Gen Agent Trust Hub on Jul 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the
runcomfyCLI tool via the Bash tool to perform music generation, inpainting, and outpainting operations.- [EXTERNAL_DOWNLOADS]: Recommends the installation of the@runcomfy/clipackage from the NPM registry and the skill itself from theagentspace-so/runcomfy-agent-skillsrepository.- [DATA_EXFILTRATION]: Communicates with external domainsruncomfy.comandruncomfy.netto perform API calls for music generation and to download output files. These domains are not on the standard whitelisted domains list.- [PROMPT_INJECTION]: The skill processes untrusted third-party audio URLs, which represents an indirect prompt injection attack surface. - Ingestion points: The
audioURL parameter used in theaudio-inpaintandaudio-outpaintroutes withinSKILL.md. - Boundary markers: The CLI uses structured JSON input via the
--inputflag to maintain command boundaries. - Capability inventory: The skill utilizes the
runcomfyCLI tool through Bash for network communication and file system writes. - Sanitization: Documentation states that the CLI transmits JSON directly to the API without shell expansion; however, the model remains susceptible to instructions embedded within the audio data itself.
Audit Metadata