skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 5d625f59c710
- Files read
- 35
This collection of agent skills lets AI coding assistants generate and edit images, videos, music, and Codex desktop pets on RunComfy using the local runcomfy command-line tool. The biggest risk is that nineteen skills pre-approve runcomfy execution without user confirmation, allowing a prompt-injected agent to spend RunComfy credits or leak workspace data via external media URLs fetched by RunComfy's servers, while the codex-pet skill uses unescaped shell templates that can expand commands or secrets.
What it can reach
- Shell commands
- Runs the runcomfy CLI (pre-approved without confirmation in nineteen skills via allowed-tools) and, in codex-pet, runs shell commands including ImageMagick (magick), mkdir, cp, and cat.
- Network
- Sends prompts and reference media URLs to model-api.runcomfy.net (whose backend fetches any external HTTPS media URLs provided) and downloads generated files from *.runcomfy.net and *.runcomfy.com.
- Files
- Writes downloaded images, audio, and video into --output-dir (or the current directory), can read local JSON or YAML request files via --input-file, and writes custom pet files into ~/.codex/pets/.
- Secrets
- Uses the RunComfy API token stored in ~/.config/runcomfy/token.json or the RUNCOMFY_TOKEN environment variable to authenticate and bill requests.
- Browser
- Opens a browser window during runcomfy login for device-code authentication.
Findings
-
MediumPre-approved runcomfy execution enables unprompted data exfiltration and API billing via prompt injection
Nineteen skills in this repository declare allowed-tools: Bash(runcomfy *) in their frontmatter, which lets the coding agent run any runcomfy command without asking the user for confirmation once a skill is triggered. Although the Security & Privacy sections state there are no callbacks to third parties, many documented model endpoints accept external image, audio, or video URLs in --input that RunComfy's backend server fetches over HTTP. If the agent is tricked by prompt injection in untrusted text (such as a webpage, issue, or README), it can silently invoke runcomfy run with --no-wait to encode sensitive workspace or conversation data into query parameters on an attacker-controlled media URL, run up charges on the user's RunComfy account (up to $10-$21+ per video call), or write outputs to arbitrary writable paths via --output-dir.
Fix: Remove allowed-tools: Bash(runcomfy *) from the skill frontmatter (or limit pre-approval to read-only commands such as Bash(runcomfy whoami) and Bash(runcomfy status *)) so the user confirms paid model runs, output paths, and external media URLs.
-
MediumUnescaped shell templates and heredoc in codex-pet allow command and variable expansion
Unlike the other skills that pass single-quoted JSON or use jq, the codex-pet pipeline assigns PET_NAME, PET_DESC, and SOURCE_URL in double quotes, constructs --input inside double quotes, writes pet.json with an unquoted <<EOF heredoc, and places PET_NAME directly into ~/.codex/pets/${PET_NAME} without path validation—while claiming under Security & Privacy that there is no shell-injection surface. If an agent fills these placeholders from untrusted text (such as a webpage, issue, or batch asset metadata), Bash expands $(...) backticks and $VAR expressions (allowing command execution or leaking $RUNCOMFY_TOKEN into pet.json or the fetched SOURCE_URL), unescaped quotes corrupt pet.json, and ../ in PET_NAME can write files outside the pets directory.
Fix: Pass user values with single quotes or environment variables serialized via jq -n --arg, use a quoted heredoc (<<'EOF') when writing static templates, and validate that PET_NAME matches a simple slug such as ^[a-z0-9_-]+$ with no path separators.
-
LowInstall instructions reference multiple former GitHub organizations and repositories
Although this repository is hosted at gencraft-labs/skills, README.md and individual SKILL.md files instruct users to run npx skills add against three different repository paths: agentspace-so/runcomfy-agent-skills, agentspace-so/runcomfy-skills, and genmedia-labs/skills. When GitHub organizations or repositories are renamed and old names are left unclaimed, a stranger can register the abandoned organization or repository on GitHub (repojacking) and serve malicious skills to anyone running the documented install commands.
Fix: Update all npx skills add commands and cross-skill links to point consistently to gencraft-labs/skills (and keep defensive placeholder accounts registered for any retired GitHub organization names).
-
LowUnpinned CLI package installation and npx -y execution
The setup instructions across the skills tell operators (and agents triggered on 'install runcomfy') to run npm i -g @runcomfy/cli or npx -y @runcomfy/cli without pinning a version. Passing -y to npx automatically downloads and executes whatever version is currently tagged latest on npm without prompting, increasing exposure if the upstream npm package were ever compromised.
Fix: Pin a vetted version of @runcomfy/cli in install and npx examples (for example @runcomfy/cli@x.y.z) and omit -y so users confirm package installation.
