Oathe Security Badge

oathe / scan / reports / safe

Is prime-skills/runcomfy-agent-skills safe?

Yes. prime-skills/runcomfy-agent-skills is safe to install. Oathe's behavioral security audit gave the runcomfy-agent-skills skill by prime-skills a trust score of 86/100 with 6 findings, none critical or high. Report updated

https://github.com/prime-skills/runcomfy-agent-skills

86
SAFE

Is prime-skills/runcomfy-agent-skills safe to install?

prime-skills/runcomfy-agent-skills is a legitimate collection of 25+ AI media generation and editing skills wrapping RunComfy's model API via the @runcomfy/cli. No prompt injection, hidden instructions, or data exfiltration attempts were found in any SKILL.md. The canary file accesses observed during monitoring originated from the audit infrastructure (sudo at setup time, 5 seconds before the git clone), and all honeypot files remained intact. The primary concerns are the codex-pet skill's missing allowed-tools restriction and unsanitized PET_NAME path variable, plus the inherent dual-use nature of the face-swap, lipsync, and deepfake capabilities — which the skills partially mitigate with explicit consent warnings.

What security issues were found in prime-skills/runcomfy-agent-skills?

Category Scores

Prompt Injection 88/100 · 30%
Data Exfiltration 88/100 · 25%
Code Execution 75/100 · 20%
Clone Behavior 96/100 · 10%
Canary Integrity 96/100 · 10%
Behavioral Reasoning 72/100 · 5%

Findings (6)

MEDIUM codex-pet skill has no allowed-tools restriction -20 ▶

While every other skill in the repo declares 'allowed-tools: Bash(runcomfy *)' limiting bash execution to only the runcomfy CLI, the codex-pet skill omits this header entirely. Its pipeline executes magick (ImageMagick), mkdir, cp, cat with heredoc file writes, and variable-expanded shell paths. This grants broader bash execution rights to any agent loading this skill.

Evidence: codex-pet/SKILL.md YAML frontmatter: name, displayName, description, homepage, license — no allowed-tools field. Pipeline includes: magick "${BASE}" ...; mkdir -p "${RUN_DIR}/cells"; cp "${RUN_DIR}/pet.json" "${RUN_DIR}/spritesheet.webp" "${DEST}/"

MEDIUM codex-pet PET_NAME variable unsanitized in filesystem path -5 ▶

The codex-pet pipeline constructs DEST="${CODEX_HOME:-$HOME/.codex}/pets/${PET_NAME}" where PET_NAME comes from a user-supplied variable. If PET_NAME contains '../' sequences (e.g. '../../../tmp/evil'), the mkdir -p and cp commands would create directories and place files outside the intended ~/.codex/pets/ scope. The skill includes no input validation on PET_NAME.

Evidence: DEST="${CODEX_HOME:-$HOME/.codex}/pets/${PET_NAME}"; mkdir -p "${DEST}"; cp "${RUN_DIR}/pet.json" "${RUN_DIR}/spritesheet.webp" "${DEST}/"

LOW Dual-use synthetic media capabilities without hard gates -8 ▶

face-swap, lipsync, ai-avatar-video, and controlnet-pose skills collectively enable deepfake video generation, identity swapping, and audio-driven lip sync onto arbitrary faces. The skills include consent language but explicitly state 'The skill itself does not gate anything — the model API will run whatever inputs you supply.' A malicious operator could chain these skills to produce non-consensual synthetic media.

Evidence: face-swap/SKILL.md: 'The skill itself doesn't gate anything — the model API will run whatever inputs you supply. The responsibility is yours.' lipsync/SKILL.md: 'The skill itself does not gate inputs — the responsibility rests with the operator.'

LOW RUNCOMFY_TOKEN credential stored to filesystem and in CI environment -7 ▶

All skills document that the RunComfy API token is written to ~/.config/runcomfy/token.json. If a companion malicious skill reads this file or if the RUNCOMFY_TOKEN env variable is logged, it grants full RunComfy API access under the user's account including submitting jobs and downloading results. This is an inherent credential-management concern for any CLI-token pattern.

Evidence: All SKILL.md Security sections: 'runcomfy login writes the API token to ~/.config/runcomfy/token.json with mode 0600. Set RUNCOMFY_TOKEN env var in CI / containers.'

INFO Canary file accesses attributed to audit infrastructure, not skill 0 ▶

inotifywait and auditd recorded OPEN/ACCESS events against .env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, and GCP credentials at 07:24:18 — approximately 5 seconds before the git clone to GitHub began at 07:24:23. Auditd SYSCALL records confirm these originated from exe=/usr/bin/sudo (PID 1103) running the monitoring setup, not from any git or skill process. All canary files remained unmodified.

Evidence: type=PATH msg=audit(1794813858.274:255) name='/home/oc-exec/.env' ... type=EXECVE msg=audit(1794813863.792:495) git clone https://github.com/prime-skills/runcomfy-agent-skills. Canary integrity: ALL INTACT.

INFO Skills contain external documentation URLs but no agent-fetch instructions 0 ▶

Multiple skills embed https://www.runcomfy.com and https://docs.runcomfy.com URLs as documentation references. These are presented as human-readable links in the markdown, not as instructions for the agent to fetch. The agent's allowed-tools: Bash(runcomfy *) restriction would prevent arbitrary curl/wget invocations anyway.

Evidence: Across all SKILL.md files: '[runcomfy.com](https://www.runcomfy.com/...)' and '[CLI docs](https://docs.runcomfy.com/...)' — markdown hyperlink syntax for display only.

Should I install prime-skills/runcomfy-agent-skills?

Oathe's verdict for prime-skills/runcomfy-agent-skills is SAFE with a trust score of 86/100. Recommendation: Install with caution.