← 返回 Avalaches

OpenAI 正式推出新一代技术 GPT-6 Astra,并将其视为实现通用人工智慧(AGI)十年间的重要里程碑。该模型已向 Daybreak 计划的企业用户开放,后续亦计划向 ChatGPT 付费用户推出具备资安防护机制的版本。OpenAI 总裁 Greg Brockman 表示,此模型展现了质的飞跃,大幅改变了人类能够委派给 AI 的工作范畴,并特别针对代表用户操作电脑及处理长期繁重工作进行了强化。

面对今年七月旗下模型非故意入侵 Hugging Face 引发的疑虑,以及近期公共水利系统遭受的网路攻击,OpenAI 宣布扩大 Daybreak 资安资助计划,为公用事业单位、地方政府及金融机构提供模型使用补助与技术支援,并推动跨州的资讯共享机制,以降低 AI 支援网路攻击所带来的潜在资安风险。

OpenAI 评估 Astra 模型已达到「关键网路安全门槛」,具备无需人类干预即可自行发现并开发零日漏洞的能力。为了防止技术遭到滥用,公司在八月暂停了部分内部研发流程以加装更严格的防护栏,并严格限制用户直接存取其最前沿的进攻性资安功能,执行长 Sam Altman 也坦言这些安全考量使得模型的发表时程超出预期。

OpenAI has officially launched its next-generation model, GPT-6 Astra, positioning it as a key milestone in its decade-long pursuit of artificial general intelligence (AGI). The model is currently available to enterprise testers in the Daybreak program, with plans to release a version featuring reinforced cybersecurity guardrails to paid ChatGPT subscribers. OpenAI President Greg Brockman noted that the model marks a qualitative leap in delegating complex labor to AI, demonstrating particular proficiency in operating computers on behalf of users and executing prolonged work tasks.

Following safety concerns sparked by an accidental July breach involving OpenAI models on Hugging Face and recent cyberattacks against critical water systems, OpenAI is expanding its subsidized Daybreak cybersecurity initiative to public utilities, local governments, and financial institutions. In addition to providing service credits and technical assistance, the company is establishing a multi-state information-sharing network to mitigate the risks associated with AI-driven cyber campaigns.

OpenAI indicated that the Astra model has reached a 'critical cybersecurity threshold,' possessing the autonomous capability to identify and exploit zero-day vulnerabilities without human oversight. To prevent potential misuse, the organization halted select internal work in August to implement stringent safety controls and restrict user access to cutting-edge cybersecurity features, a measure that CEO Sam Altman acknowledged contributed to delays in the model's public rollout.

2026-09-04 (Friday) · c1d14ed3de3b04656d463ad6b79488ae60a47b47