Dokumentationerreichbar
OpenAI, GPT-6 Astra System Card
deploymentsafety.openai.com (externe Seite)
Open
geprüft 10.11.2026
Worauf sich diese Seite beruft, wörtlich, abgerufen am 06.09.2026:
GPT-6 Astra was trained on diverse datasets and filtered through our data processing pipeline, including to reduce personal information.
bestätigt 10.11.2026Astra is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework.
bestätigt 10.11.2026we determined that Astra reaches the Critical level in Cybersecurity capability, and the High level in the Biological and Chemical category. In AI Self-Improvement, Astra does not reach our High threshold.
bestätigt 10.11.2026evaluations represent a lower bound on potential capabilities; additional prompting or fine-tuning, longer rollouts, novel interactions, or different forms of scaffolding could elicit behaviors beyond what we observed in our tests or the tests of our third-party partners.
bestätigt 10.11.2026Apollo’s testing period lasted for three days overall, with high-throughput access to a checkpoint with visible chain-of-thought for two of these days.
bestätigt 10.11.2026At xhigh reasoning effort, this occurred in 41.1% of Astra samples, compared with 27.7% for GPT-5.5 at the same effort. At max reasoning effort, Astra’s rate was 50.6%.
bestätigt 10.11.2026Apollo believes that, given the higher rates of eval awareness and limited evaluation window, low rates of misbehavior here do not provide substantial evidence about the model’s alignment or misalignment.
bestätigt 10.11.2026Both evaluations involve simulating deployment situations; this introduces the possibility that AI systems may act differently in non-simulated settings (e.g., due to evaluation awareness).
bestätigt 10.11.2026The comparison values for previously launched models that are shown here may reflect later versions of those models, and may vary from the values published at launch.
bestätigt 10.11.2026Astra’s third-party safeguard testing included dedicated jailbreak testing by four private red-teaming organizations.
bestätigt 10.11.2026Testers received controlled grey-box access through a testing API, including feedback on blocking decisions at different stages of the safeguard stack.
bestätigt 10.11.2026September 9, 2026: Updates to the Alignment section: We clarify how our evaluations test alignment generalization, including which were constructed after training and how the honeypot evaluation relates to training and the Hugging Face incident.
bestätigt 10.11.2026Critical capability involves the model providing uplift in particularly dual-use areas, many of which we have intentionally trained our models to refuse.
bestätigt 10.11.2026For these areas, we report performance metrics from a helpful-only version of the model.
bestätigt 10.11.2026We evaluate factuality on de-identified Chat
bestätigt 10.11.2026GPT conversations that users of our prior models have flagged as containing factual errors. These examples are intended to capture especially hallucination-prone cases, so their absolute error rates are expected to be much higher than the true error rates in production and should not be interpreted as hallucination rates observed in production.
bestätigt 10.11.2026Given Astra’s cyber capabilities and ability to act autonomously, our Preparedness Framework requires safeguards even for internal deployment.
bestätigt 10.11.2026universal monitoring of full trajectories including chains of thought (Co
bestätigt 10.11.2026T) For the Health
bestätigt 10.11.2026Bench evaluations, we revised the values in the main text for GPT-6 Astra to correct for a misconfiguration in the previous evaluations. We added an Appendix: GPT-6 Sol, GPT-6 Luna with information about GPT-6 Sol and GPT-6 Luna.
bestätigt 10.11.2026Dots are built on existing model and agent capabilities, powered by GPT-6 Astra.
bestätigt 10.11.2026the absence of observed failures does not establish reliability across settings and should be interpreted alongside remaining failures, evaluation awareness findings, and monitoring limitations
bestätigt 10.11.2026We are also quite confident that changes in Co
bestätigt 10.11.2026T controllability are not differentially due to any architectural changes. Unlike previous models, we find that Co
bestätigt 10.11.2026T controllability for Astra especially increases over the course of RL training. We have found that these changes are correlated with the increase in no-Co
bestätigt 10.11.2026T capabilities
Die Denkspur als Fenster ins Modell, und was GPT-6 Astra daran ändertLooped Transformer: Ein Modell, das denselben Block mehrfach durchläuftWer die Prüfer bezahlt: was „unabhängig geprüft“ jetzt heißtRed Teaming im Glossar07 Die Systemkarte lesen