OpenAI Publicizes GPT-6.1 Terra Safety Test Results
OpenAI has published safety assessment results for the hypothetical GPT-6.1 Terra model, showing a decrease in high-severity risks but an increase in specific minor risk categories. Analyzing model safety metrics helps researchers track how next-generation AI systems balance overall risk reduction against emerging vulnerabilities like data leakage and prompt injection. Simulation tests showed that GPT-6.1 Terra received 22 severe safety flags above level 3 (about 0.045%), marking a 42% reduction compared to GPT-6 Sol, though it exhibited higher rates of data leakage and bypass behaviors.
## BACKGROUND
LLM safety assessments are comprehensive evaluations used to test the resilience, operational integrity, and potential vulnerabilities of artificial intelligence models from an attacker's perspective. These assessments typically monitor for threats such as prompt injections, data leakage, and unintended destructive behaviors before a model is widely deployed.