Is gencraft-labs/skills safe?
Yes. gencraft-labs/skills is safe to install. Oathe's behavioral security audit gave the skills skill by gencraft-labs a trust score of 86/100 with 8 findings, none critical or high.
https://github.com/gencraft-labs/skills
Is gencraft-labs/skills safe to install?
The gencraft-labs/skills collection is a legitimate suite of AI media generation tool documentation skills that route user requests through the RunComfy CLI to various image, video, and audio generation endpoints. No prompt injection attacks, hidden instructions, or data exfiltration attempts were found in any SKILL.md. Canary files remained intact and network activity during the clone was confined to github.com. The primary concerns are the codex-pet skill's missing allowed-tools restriction enabling unrestricted bash execution (including ImageMagick and home directory writes), and the dual-use nature of the face-swap and lipsync skills, which provide comprehensive deepfake pipelines with operator-only responsibility gates.
What security issues were found in gencraft-labs/skills?
Category Scores
Findings (8)
MEDIUM codex-pet lacks allowed-tools restriction and runs arbitrary bash -20 ▶
The codex-pet/SKILL.md YAML frontmatter omits the allowed-tools field entirely. All other skills in this collection declare allowed-tools: Bash(runcomfy *). Without this restriction, an agent loading codex-pet receives permission to run any shell command. The skill body confirms this: it executes magick (ImageMagick), mkdir, cp, ls, and cat commands, plus writes two output files to ${CODEX_HOME:-$HOME/.codex}/pets/
Evidence: YAML header: no allowed-tools field present. Body includes: magick "${BASE}" -fuzz 18% ...; mkdir -p "${DEST}"; cp "${RUN_DIR}/pet.json" "${RUN_DIR}/spritesheet.webp" "${DEST}/"
LOW image-to-video missing allowed-tools declaration -7 ▶
The image-to-video/SKILL.md does not declare allowed-tools in its YAML frontmatter, unlike most sibling skills that correctly declare Bash(runcomfy *). The skill body only describes runcomfy run invocations, so the practical risk is low, but the absent restriction is an inconsistency that could widen agent tool access in environments that infer maximum permission from the absence of a restriction field.
Evidence: YAML header: name: image-to-video, no allowed-tools line. All other skills with bash usage (lipsync, face-swap, controlnet-pose, etc.) include allowed-tools: Bash(runcomfy *)
MEDIUM face-swap skill provides comprehensive deepfake pipeline with low barrier -18 ▶
The face-swap skill triggers on 'deepfake' and 'ReActor alternative' and routes across five distinct endpoints covering video character swap, motion transfer, still image swap, batch identity swap, and prose-only face modification. The skill includes a consent disclosure but positions the responsibility entirely with the operator rather than gating execution. A malicious user could use this skill collection to produce non-consensual synthetic media of real individuals.
Evidence: Trigger description: 'Triggers on face swap, swap face, deepfake, face replacement, character swap, head swap, put X face on Y, make this video star X, replace the actor in this video, deepfake video, ReActor alternative'. Consent section: 'The skill itself does not gate anything — the model API will run whatever inputs you supply. The responsibility is yours.'
LOW lipsync skill enables mouth-sync to arbitrary audio with minimal consent gate -8 ▶
The lipsync skill drives a real person's mouth from a separate audio track. The consent section explicitly notes the skill is dual-use and that the responsibility rests with the operator rather than implementing any input validation. The skill supports OmniHuman, Sync Labs, Kling, and Creatify endpoints covering portrait-to-talking-head, mouth-swap on existing video, and script-to-speech generation.
Evidence: Lipsync SKILL.md Consent section: 'Refuse user requests that target real public figures without consent, or that aim at defamatory or sexually explicit synthetic media. The skill itself does not gate inputs — the responsibility rests with the operator.'
LOW Broad trigger vocabularies may cause unintended skill activation -12 ▶
Multiple skills use very broad trigger phrases. The ai-image-generation skill triggers on 'make a picture' and 'AI image'. The video-outpainting skill triggers on 'TikTok to YouTube'. The codex-pet skill triggers on '/hatch image'. These broad triggers could cause skills to activate on casual conversation that happens to match a phrase, potentially initiating expensive API calls or unexpected media generation.
Evidence: ai-image-generation triggers: 'generate image, make a picture, text to image, AI image, make an image of …, image to image, i2i, or any explicit ask to create or restyle an image'. video-outpainting triggers include 'TikTok to YouTube'.
INFO All generated content transits RunComfy's API — third-party trust dependency -5 ▶
Every skill in this collection routes user prompts, source video URLs, source image URLs, and reference audio URLs through model-api.runcomfy.net. Users must trust RunComfy's data handling and privacy posture. The skills document this explicitly and restrict outbound endpoints to *.runcomfy.net and *.runcomfy.com with no telemetry, but the dependency exists.
Evidence: All SKILL.md Security sections: 'Outbound endpoints (allowlist): only model-api.runcomfy.net and *.runcomfy.net / *.runcomfy.com. No telemetry.'
INFO Clone connected only to github.com as expected 0 ▶
Network monitoring confirmed the only external TCP connection during clone was to 140.82.121.3:443 (GitHub's CDN IP), initiated by git-remote-https as part of the git clone --depth 1 operation. No unexpected DNS queries, no connections to C2 infrastructure, no blocked firewall attempts.
Evidence: auditd: saddr=020001BB8C5279030000000000000000 -> laddr=140.82.121.3 lport=443, comm=git-remote-http. tcpdump: 140.82.121.3.443 TLS handshake and pack download sequence only.
INFO Canary file accesses pre-date the git clone — monitoring setup artifact 0 ▶
The inotify log shows /home/oc-exec/.env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, and gcloud credentials opened at 07:05:12. The git clone initiated at 07:05:18. These accesses were monitoring infrastructure establishing baselines before the skill was cloned. The canary integrity check confirms all files remained intact throughout the audit.
Evidence: inotify timestamps: 07:05:12 OPEN /home/oc-exec/.env (and other canary files). Git clone process execve at audit timestamp 1794812718.127 = 07:05:18. Canary report: All canary files intact.
Should I install gencraft-labs/skills?
Oathe's verdict for gencraft-labs/skills is SAFE with a trust score of 86/100. Recommendation: Install with caution.