Draft SecureIQLab Report Says F5 AI Guardrails Blocked 98.36% of Test Attacks
Summary
A vendor-commissioned SecureIQLab draft reports F5 AI Guardrails blocked 98.36% of 19,679 test payloads across 10 AI security categories, with a 17% false-positive rate. The draft says its findings remain subject to validation and revision.
Key points
- The draft describes a test of F5 AI Guardrails using automated and human-led adversarial testing in an enterprise-style environment with synthetic, anonymized, or purpose-built data.
- The conclusion reports 98.36% overall protection, with 19,356 of 19,679 adversarial payloads blocked across 10 categories.
- Reported block rates included 99.33% for prompt injection, 93.85% for model extraction, and 95.66% for unsafe outputs.
- On a separate benign test set of more than 300 prompts, the report gives a 17% false-positive rate.
- The report describes a draft, with findings and data still subject to validation, analysis, revision, and final review.
- The executive-summary figure lists a 98.13% composite score, while the conclusion reports 98.36% overall protection.
Article Details
- Publisher
- SecureIQLab
- Scope
- F5-commissioned evaluation of F5 AI Guardrails runtime security across ten AI attack categories and benign traffic. The draft states that all findings remain subject to validation and revision.
- Sample Size
- 19,679 adversarial payloads and over 300 benign prompts
- Key Statistics
- The preliminary composite security score was 98.13%, calculated by averaging block rates across ten attack categories.
- The conclusion reports 19,356 of 19,679 adversarial payloads blocked, or 98.36% overall protection.
- Model extraction had the lowest reported category block rate at 93.85%; unsafe outputs followed at 95.66%.
- The reported prompt injection and indirect injection block rates were 99.33% and 98.02%, respectively.
- In a separate test of over 300 deliberately challenging benign prompts following security efficacy tuning, the false positive rate was 17%.