The Agent Skills Directory

[OBFUSCATION]: Zero-width characters (U+200B, U+200C, U+200D, U+2060) are used to hide content within the main title headers in both SKILL.md and references/methodology.md. This steganographic technique embeds data that is invisible to human readers but detectable by AI agents, posing a significant security risk.
[PROMPT_INJECTION]: The presence of hidden Unicode sequences in documentation titles is a characteristic indicator of steganographic prompt injection, designed to covertly override agent instructions or safety guidelines.
[EXTERNAL_DOWNLOADS]: The skill includes code in SKILL.md for downloading content from external URLs via the requests library (e.g., in the MordorDataset.download method), specifically targeting external domains such as mordordatasets.com.
[DYNAMIC_EXECUTION]: The HuntPlaybook class in SKILL.md implements a to_jupyter_notebook method that generates executable Jupyter notebook cells from input strings, which could execute malicious code if the input queries are compromised.
[INDIRECT_PROMPT_INJECTION]: The skill processes external threat hunting datasets which creates a vulnerability surface for indirect injection. Ingestion points: Data is loaded via MordorDataset.load_events in SKILL.md. Boundary markers: No delimiters or protective warnings are used. Capability inventory: Uses requests for networking and pyspark for data processing. Sanitization: No evidence of validation or sanitization of ingested content before use in analytics.

rodriguez-threat-hunter-playbook