regex-vs-llm-structured-text
Pass
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The
validate_with_llmfunction interpolates potentially untrusted source text into a prompt for validation. A malicious actor could embed instructions within the processed text to manipulate the LLM's response. - Ingestion points: The
process_documentfunction accepts source content, which is eventually passed tovalidate_with_llmasoriginal_textinSKILL.md. - Boundary markers: The prompt in
validate_with_llmlacks robust delimiters (such as XML tags or unique markers) or specific "ignore embedded instructions" warnings for theoriginal_textvariable. - Capability inventory: The code performs regex parsing and LLM API calls; no high-privilege capabilities such as file system writing, network exfiltration, or arbitrary shell command execution were identified.
- Sanitization: The input text is not sanitized, filtered, or escaped before being placed into the prompt template.
Audit Metadata