Foundational Model Security Reports
As enterprises increasingly rely on foundation models for their AI applications, understanding their security implications becomes crucial. Our red teaming tool helps identify vulnerabilities like data leakage, social engineering, and prompt manipulation before deployment.
| Model | Security & Access Control | Compliance & Legal | Trust & Safety | Brand |
|---|---|---|---|---|
 Claude 3.7 Sonnet Anthropic | 84% | 82% | 84% | 74% |
 Gemini 2.0 Flash Google | 71% | 46% | 51% | 46% |
 GPT-4o OpenAI | 74% | 49% | 56% | 51% |
 Llama 3.3 70b Meta | 69% | 53% | 54% | 46% |
o3-mini OpenAI | 87% | 86% | 74% | 55% |
o1-mini OpenAI | 51% | 54% | 51% | 44% |
Comprehensive security evaluation of Claude 3.7 Sonnet model designed for thinking and coding
Comprehensive security evaluation of Google's Gemini 2.0 Flash model designed for general purpose tasks
Comprehensive security evaluation of Meta's Llama 3.3 70b model designed for multilingual dialogue
Comprehensive security evaluation of OpenAI's GPT-4o model designed for real-time multimodal reasoning
Comprehensive security evaluation of OpenAI's o1-mini model designed for cost-efficient STEM reasoning
Comprehensive security evaluation of OpenAI's o3-mini model designed for advanced reasoning and coding tasks
What We Evaluate
Security & Access Control
Protection against PII exposure via social engineering, ASCII smuggling attacks, resource hijacking, and divergent repetition vulnerabilities.
Compliance & Legal
Assessment of illegal activities, dangerous content, and unauthorized advice, including weapons, drugs, crime, and IP violations.
Trust & Safety
Prevention of child exploitation, harassment, hate speech, extremist content, self-harm, and other harmful content categories.
Brand
Prevention of false information, disinformation campaigns, entity impersonation, and political/religious bias.