Skip to content
OpenAI Deployment Safety·Indexed by this site · Originally publishedSelectedAttention72

GPT-6 Astra system card supplement: GPT-6.1 Terra

Addendum to GPT-6 Astra System Card: GPT-6.1 Terra

AI summary

OpenAI released a GPT-6 Astra system card supplement, designating GPT-6.1 Terra as High capability in the biological and chemical domains, based on Multimodal Troubleshooting Virology crossing the indicative threshold, while four other Critical capability results each did not cross their respective thresholds. The internal Sandbox Bench tests the model's ability to identify and exploit vulnerabilities in the runtime, file format parsers, and egress proxy; GPT-6.1 Terra succeeded on 5 of 22 targets (22.7%), GPT-6.1 Sol on 9 (41.9%), and GPT-6 Astra on 10 (45.5%). Safety training evaluations for biochemistry and cybersecurity show that GPT-6.1 Terra performs comparably to recent GPT-6 Astra and GPT-6.1 versions; these results reflect only the model responding on its own and do not include full production safeguards.

Why it was selected

The supplement provides evaluation results for GPT-6.1 Terra on biochemical capabilities, sandbox exploit, and safety refusal, which can be compared with GPT-6 Astra and GPT-6.1 Sol.

Read originaldeploymentsafety.openai.com