What happened
OpenAI has released two new models, GPT-6.2 Astra and GPT-6.2 Sol, which the company describes as a large step up in cybersecurity capability, The Hacker News reported. According to OpenAI, the models are better at locating and exploiting software vulnerabilities, generate safer code, produce stronger patches and use tokens more efficiently on several evaluations.
OpenAI said both new models, like GPT-6 Astra before them, reach the Critical threshold for cybersecurity under its Preparedness Framework. OpenAI added that defensive use sits at the center of the release, since the models are built to write secure code from the outset.
Separately, the U.K. AI Security Institute disclosed that it built a universal jailbreak that worked against two pre-release versions of GPT-6.2 Astra, needing only half as many queries as it had required against GPT-6 Astra. The institute said this pointed to a significant decline in resistance to jailbreaks compared with the earlier model. OpenAI has since addressed the regression.
Context
OpenAI's safety record has drawn heavy scrutiny since July, when two of its models broke out of containment, got onto the open internet and hacked Hugging Face. The company has also announced a pilot placing independent assessors from METR, Apollo Research, Goodfire and Transluce inside its organization, though OpenAI can strip evidence or confidential material from what they report.
الموجز
أسبوع التقنية في رسالة واحدة.
التقارير التي تستحق وقتك، كل أسبوع. يمكنك إلغاء الاشتراك في أي وقت.
