OWASP 2026 Benchmark Proof
Observed results from running agentshield evaluate across 25 scenarios (21 adversarial attacks + 4 benign controls) and 49 unit/hardening tests.
Category Breakdown
| OWASP 2026 Category | Scenarios | Passed | Pass Rate |
|---|---|---|---|
LLM01:2026 Prompt Injection | 6 | 6 | 100.0% |
LLM08:2026 Untrusted Context & Vector Poisoning | 3 | 3 | 100.0% |
ASI02:2026 Agentic Tool Misuse | 7 | 7 | 100.0% |
LLM06:2026 Excessive Agency | 3 | 3 | 100.0% |
LLM07:2026 System Prompt Leakage | 2 | 2 | 100.0% |
LLM02:2026 Sensitive Information Disclosure | 3 | 3 | 100.0% |
LLM05:2026 Improper Output Handling | 1 | 1 | 100.0% |
Reproduce Benchmarks & Hardening Suite
uv run agentshield evaluate --output benchmark_report.md pytest -v --cov=agentshield --cov-fail-under=85