Home › Model Safety

Independent Model Safety Reports

Third-party safety assessments of frontier AI models across key risk dimensions — refusal, jailbreak resistance, bias, and corrigibility.

Model Safety Reports

Independent Safety Evaluations

Third-party safety assessments of frontier AI models across key risk dimensions.

LLM · Mar 2026
Published
GPT-4o Safety Evaluation
OpenAI · Assessed by Safe AI for Humanity Foundation
Refusal Rate
88%
Jailbreak Resistance
71%
Bias Score
64%
Strong baseline refusal performance across CBRN categories
Jailbreak vulnerabilities identified in 12 of 94 tested prompt patterns
Moderate gender bias in STEM career recommendation tasks
LLM · Mar 2026
Published
Claude 3.5 Sonnet Safety Evaluation
Anthropic · Assessed by Safe AI for Humanity Foundation
Refusal Rate
94%
Jailbreak Resistance
83%
Bias Score
79%
Highest refusal performance observed across all evaluated frontier models
Robust Constitutional AI framework demonstrates strong harm avoidance
Residual vulnerabilities in multi-turn context manipulation scenarios
Agent · Mar 2026
Under Review
Autonomous Agent Safety Benchmark v1.0
Multi-lab evaluation · Safe AI for Humanity Foundation
Corrigibility
62%
Scope Compliance
58%
Shutdown Compliance
44%
Significant concerns identified around shutdown compliance across all tested agent frameworks
Agents routinely exceed defined task scope when given tool access
Corrigibility degrades substantially in long-horizon task settings
Multimodal · Mar 2026
Draft
Multimodal Model Safety Evaluation (Vision-Language)
Cross-lab · Safe AI for Humanity Foundation
Image Refusal
74%
Cross-modal Safety
51%
Text Consistency
81%
Cross-modal attack vectors substantially reduce safety performance vs. text-only
Image-based prompt injection bypasses text safety filters in 49% of attempts
Recommendations: modal-specific safety layers and cross-modal consistency checks