Agent Skill Malware: Malicious vs Benign OpenClaw Skill Files

350 deduplicated OpenClaw agent skill files (SKILL.md) from February 2026, comprising 127 malicious and 223 benign samples. The dataset was created by yoonholee and last updated on Hugging Face on 2026-02-26. Malicious samples were extracted from real campaigns targeting ClawHub and use social engineering in markdown instructions to trick agents into running malware like AMOS (Atomic macOS Stealer).

Use Cases

• Train binary classifiers to distinguish malicious from benign agent skills based on markdown content.
• Analyze social engineering tactics and malware delivery patterns within agent skill instructions.
• Benchmark security tools for detecting malicious code or instructions in AI agent ecosystems.

Strengths

• Contains 350 samples, a specific split of 127 malicious and 223 benign files.
• Samples are deduplicated by content hash, reducing redundancy.
• Malicious samples are sourced from real campaigns in February 2026, providing contemporary examples.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Malicious skills from openclaw/skills GitHub archive; benign skills source not specified.

Collection Method

Extracted from malicious campaigns targeting ClawHub and deduplicated by content hash.

Time Range

February 2026

Freshness

Last updated 2026-02-26 17:57:58; freshness should be verified.