roth-detection-engineering
Warn
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: MEDIUMOBFUSCATIONPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [OBFUSCATION]: Hidden zero-width characters are embedded in the headers of SKILL.md and references/sigma_reference.md. These characters form an invisible bitstream designed to be processed by the LLM tokenizer while remaining hidden from human reviewers.- [PROMPT_INJECTION]: The steganographic content in the skill's documentation is a known vector for prompt injection. This technique attempts to override the agent's system instructions and safety filters by providing hidden commands that the model follows.- [INDIRECT_PROMPT_INJECTION]: The skill provides tools for parsing and converting Sigma rules from YAML format, creating an attack surface for indirect prompt injection.
- Ingestion points: The SigmaConverter and RuleQualityAnalyzer components in SKILL.md and scripts/sigma_converter.py ingest external YAML data.
- Boundary markers: The processing logic lacks delimiters or boundary markers to distinguish between data and instructions within the rule files.
- Capability inventory: The skill includes Python scripts capable of local file reading and complex string manipulation.
- Sanitization: While the skill uses yaml.safe_load for parsing, it does not validate or sanitize the content of the rule fields (e.g., descriptions, authors) before processing.
Audit Metadata