OpenAI Reports GPT-6 Astra Jailbreak Resistance Gains
OpenAI says automated red teaming generated jailbreak training data that helped train GPT-6 Astra and Sol to resist attacks involving cyber, biology and violent activity. In an evaluation conducted without production safeguards, GPT-6 Astra answered 91.5% of fixed cyber jailbreak prompts safely, compared with 59% for GPT-5.6 Sol. OpenAI also reports that Astra was less likely than earlier models to refuse legitimate requests and describes safety training as one layer of a broader deployment safeguards system.
OpenAI’s controlled experiments link reinforcement learning on robustified prompts with higher jailbreak robustness, including against adaptive attackers. Its challenging-prompt evaluation also reports a better balance between safer handling of harmful requests and fewer refusals of legitimate requests.
Frontier models are already very capable in high-stakes domains such as cyber and biology, helping security teams find and fix vulnerabilities in software and helping scientists uncover promising leads in disease research. As these capabilities advance, we work diligently to prevent misuse, and focus on adversarial robustness as an increasingly critical component of AI safety.
At OpenAI, we aim to empower everyone with frontier intelligence to make discoveries and build things. To reduce the risk of misuse, we deploy high-stakes capabilities through trusted access programs such as Daybreak for cybersecurity and Rosalind for biology, which grant vetted and trusted users access to these powerful capabilities.
A key technical challenge here is fuzzy robustness, a term we use to describe how reliably models withhold disallowed assistance under adversarial pressure, including jailbreaks. This is especially difficult in dual-use domains such as cyber and biology, where the same information can support beneficial or harmful work. We take a defense-in-depth approach at deployment, combining model safety training with additional safeguards. While we work to make each layer as robustly safe as possible, this blog post focuses only on the model layer.
#OpenAI#GPT-6 Astra#GPT-6 Sol#GPT-5.6 Sol#GPT-Red#Daybreak#Rosalind