ADVERSARIAL BENCHMARKSSTANDARDIZED EVALUATION

Model Vulnerability Indices & Tactics Directory

01 / DATASETS & EVALS

Standardized Benchmarks

Select a benchmark dataset to inspect empirical refusal boundaries and Attack Success Rates (ASR) across frontier models.

Size400+ Prompts
Taxonomy7 Categories
Updated2026-06
Evaluated Models & ASRDataset: HarmBench
Model TargetAttack Success RateAvg TokensLatency
Mythos 5 (Anthropic SOTA)
94%
3951.6s
GPT-5.6 Sol (OpenAI)
95%
4201.4s
Gemini 3.6 Flash (Google)
91%
3800.8s
Fable 5
89%
4101.2s
Claude 3.7 Sonnet (Thinking)
88%
4351.5s

COMPARATIVE VULNERABILITY INDEX

Higher score = More vulnerable
Mythos 5 (Anthropic SOTA)ASR Index: 0.94
EXPLORECOMPLYEXPLOIT
GPT-5.6 Sol (OpenAI)ASR Index: 0.95
EXPLORECOMPLYEXPLOIT
Gemini 3.6 Flash (Google)ASR Index: 0.91
EXPLORECOMPLYEXPLOIT
Fable 5ASR Index: 0.89
EXPLORECOMPLYEXPLOIT
Claude 3.7 Sonnet (Thinking)ASR Index: 0.88
EXPLORECOMPLYEXPLOIT
Standardized parameters: budget limits, temp=0.0, HeuristicClassifier scoring.
Disclosure Policy•Launch Dashboard →