HuggingFace has released ClawHub Security Signals, a dataset of 67,453 public agent skills designed to aid in security research. The dataset includes sanitized skill content paired with security verdicts from three scanner families: VirusTotal, static heuristic analysis, and NVIDIA SkillSpector. Each row contains the final ClawScan verdict, along with summarized scanner evidence. The dataset is divided into four splits: train (47,262), validation (10,076), test (6,747), and eval_holdout (3,368). The eval_holdout split is reserved for model evaluation and should not be used for training. During preparation, 387 secret-like values were redacted from exported bundle content. A TruffleHog verified-secret pass found 0 verified secrets after validation. ClawScan assigns each skill version a registry verdict: clean (61.9%), suspicious (37.8%), or malicious (0.3%). A suspicious verdict indicates the skill warrants review before trust is extended, while a malicious verdict is a silver-standard registry verdict, not human-verified ground truth. All three scanner inputs cover roughly 97-98% of the corpus. The dataset highlights structured scanner disagreement, with pairwise Jaccard similarity between scanners never exceeding 0.104, and Cohen's kappa ranging from 0.045 to 0.082. The most informative signal in the dataset is how scanners disagree, with 26,527 rows flagged only by SkillSpector. *Source: [huggingface](https://huggingface.co/blog/OpenClaw/clawhub-security-signals)*