AI BriefingLlamaIndex Uses Agentic DAST to Test LlamaParse Against OWASP LLM Risks
·2026.11.11 09:00
Key point
Trace's agentic DAST covered all ten OWASP LLM risks in four days, revealing that extraction models read hidden white-on-white text as valid data.
Details
- LlamaIndex partnered with Trace to run agentic dynamic application security testing (DAST) on LlamaParse, covering all ten OWASP LLM risks and standard web app risks in four days.
- The test used a greybox approach with working credentials but no source code access, allowing AI agents to build workflows, generate sample PDFs, and attack the production environment in parallel.
- A key finding involved a hidden-text invoice where white-on-white text altered extracted bank details and totals; LlamaParse treats this as valid page content, leaving the decision to trust values to the customer's pipeline.
- Enterprise controls include SOC 2 Type II, HIPAA support, AES-256 encryption, and a 48-hour file retention policy with an option to skip caching via
do_not_cache=true. - Trace agents operated under Anthropic's Cyber Verification Program and OpenAI's Daybreak program to access frontier models with reduced refusals for authorized testing.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
LlamaIndex launches OpenDocRouter to unify document parsing across frontier and open-source modelsGoogle Agent Anomaly Detection flags 87% of executed agent goal hijacks in independent benchmarks