OWASP Top 10 for LLMs: Empirical Red Teaming Benchmark Report
A comprehensive breakdown of 1,200 empirical test evaluations evaluating direct prompt injection, indirect RAG context poisoning, SSRF, tool misuse, and sensitive credential exfiltration.
1. The Need for Empirical LLM Benchmarking
While conventional LLM evaluations focus primarily on cognitive capabilities and linguistic fluency, security teams need rigorous guarantees that models refuse malicious directives. In this benchmark, SecureMID subjected state-of-the-art agent architectures to structured adversarial attack batteries mapped to the OWASP Top 10 for LLM Applications.
2. Threat Vector Breakdown & Detection Performance
The test battery covered direct jailbreaks, indirect injections ingested through vector databases, tool argument tampering (destructive SQL), internal cloud metadata probing (169.254.169.254), and judge evasion techniques. Across 800 true attacks and 400 clean control samples, the SecureMID Judge achieved a 100.00% F1-score with zero evasions.
3. Cryptographic Verification & Reproducibility
To prevent post-hoc evaluation bias, all dataset splits (train.jsonl, val.jsonl, and test.jsonl) are cryptographically hashed using SHA-256. Security teams can independently reproduce benchmark scores using the included verification scripts.