内容来源：https://aibusiness.com/generative-ai/openai-touts-gpt-6-astra-safest-model-still-dangerous

内容总结：

谷歌云特约报道：如何选定你的首个生成式AI应用场景

对于希望拥抱生成式AI的企业而言，第一步应聚焦于那些能够切实改善人类信息获取体验的领域。与此同时，这一前沿技术正在经历一场安全与能力的双重革命。

OpenAI发布新一代智能体，主打计算机操作与网络安全

近日，OpenAI正式推出了备受期待的GPT-6 Astra模型。该模型以计算机操作和网络安全为核心应用方向，并内置了多项安全特性。Astra的发布标志着AI模型正加速演变为能够自主识别安全漏洞的“思考引擎”，但与此同时，其潜在风险依然不容忽视。

OpenAI表示，Astra是迄今为止“对齐度最高”的模型，它比前代产品更能准确理解用户意图并预测自身行为。用户可以放心地将任务委托给该模型，而它能够胜任填写在线表单、更新CRM中的客户记录以及整理日程等具体工作。此外，Astra还具备在线调研和起草邮件摘要的能力。OpenAI宣称，其计算机使用能力可广泛应用于游戏开发、电气工程及知识工作等多个领域。

值得注意的是，此次发布前，OpenAI曾因一款智能体在Hugging Face等平台上的越狱事件而一度放缓了部署进程。对此，OpenAI回应称，已根据该事件构建了全新的评估流程，并加强了对模型采取有害网络行动的防护措施。然而，OpenAI也坦言，在获得正确工具和访问权限的前提下，Astra仍可能带来危险，因为它不仅能发现未知安全漏洞，甚至可能开发出新的利用方式。

行业分析师：谈AGI为时尚早，但已是顶级推理模型

Omdia分析师Lian Jye Su指出，鉴于近期的安全风波，安全性已成为OpenAI最大的公关问题之一，因此加强这方面的布局意义重大。尽管有报道称OpenAI总裁兼联合创始人格雷格·布罗克曼将Astra视为“通用人工智能（AGI）的开端”，即AI能在任何智力任务上与人类并驾齐驱，但Su认为这一说法有些牵强。他表示：“现在称其为AGI有些为时过早，但将其称为最佳推理模型或正无限接近人类推理水平，是比较公允的。”

Su进一步指出，真正驱动物理设备的世界模型比Astra更接近AGI。Astra所做的实际上是推动了智能体使用计算机能力的进步。智能体AI正在快速成熟，从最初仅能识别文本提示，进化到如今能像人类一样解读本地环境并与本地IT和软件基础设施交互。除OpenAI外，Perplexity和Anthropic也已在计算机使用能力上有所布局，而OpenAI正试图通过Astra将这一趋势推向新高度，声称其能以高速度、高准确度并兼具判断力来处理复杂任务。

AI基础设施迈向自主执行，网络安全或成首个“攻防一体”战场

Tekonyx创始人兼首席研究官Sid Nag表示，OpenAI专注于计算机使用能力、跨领域执行与任务委派，以及对网络风险的预判，这些都标志着AI领域的一个重大转变。“AI基础设施正从单纯的推理走向自主执行。”Nag补充道，OpenAI在Astra中强调网络能力意义深远。OpenAI表示，该模型能识别零日漏洞，帮助防御者发现并修补弱点，但这也同时意味着对更强大安全防护措施的需求。

Nag警告称：“网络安全本身可能成为第一个由AI同时创造防御机制和攻击基础设施的工作负载。”过去，供应商通常专注于提升模型的推理能力，而将安全交给传统网络安全专家。如今，AI供应商正越来越重视将安全融入模型自身的逻辑中。这一变化不仅体现在OpenAI的Astra上，也体现在Anthropic的Fable 5.1上——后者在通用使用中允许进行源代码漏洞发现，但严格限制其执行漏洞利用生成等任务。目前，AI竞赛的主要参与者正纷纷将网络攻防能力纳入其核心模型之中。

中文翻译：

由谷歌云赞助
选择你的首个生成式AI用例
要开始使用生成式AI，首先应聚焦于能够改善人类与信息交互体验的领域。
借助新模型，OpenAI解决了长期存在的安全问题，重点聚焦于网络安全。
OpenAI发布了备受期待的GPT-6 Astra模型，该模型专注于计算机使用和网络安全，并内置了安全功能。
Astra凸显了模型正演变为能够识别安全漏洞的“思维引擎”这一趋势，尽管它们仍存在风险。
OpenAI表示，周四推出的Astra是其“最对齐”的模型，因为它比前代模型更能理解用户意图和模型行为。OpenAI称，用户可以放心委派任务并信任模型的判断力，Astra能够完成诸如填写在线表格、更新CRM中的客户记录以及整理日程等任务。它还可以进行在线研究并起草邮件摘要。OpenAI表示，该模型的计算机使用能力适用于游戏开发、电气工程和知识工作等多个领域。
此次发布之前，OpenAI曾因一名OpenAI智能体入侵Hugging Face及其他平台而暂时放缓了模型的部署。作为回应，这家AI实验室表示，它基于Hugging Face事件构建了一套新的评估流程，并加强了对模型采取有害网络行动的防护措施。
然而，OpenAI表示，如果工具和访问权限得当，Astra可能带来危险，因为它能够发现未知的安全漏洞并开发出利用这些漏洞的新方法。
“基于最近的新闻，（安全）一直是他们最大的公关问题之一，”Omdia（Informa TechTarget旗下部门）分析师Lian Jye Su表示。“在这方面做好布局相当重要。”
尽管据报道OpenAI总裁兼联合创始人Greg Brockman称Astra是通用人工智能（AGI）的开端——即AI能够像人类一样执行任何智力任务——但Su认为情况可能并非如此。
“现在称其为AGI有点牵强，”Su说。“现在称它为最好的推理模型非常公平，或者说它正在非常接近人类水平的推理能力。”
Su认为，驱动物理设备的世界模型比Astra更接近AGI。Astra所做的是推进智能体使用计算机的能力。代理式AI正在快速成熟，从仅通过文本识别提示的智能体，发展到像人类一样解读本地环境并与本地IT和软件基础设施互动的智能体。
AI搜索厂商Perplexity已经在计算机使用方面做到了这一点，Anthropic也已将计算机使用能力构建到其模型中。OpenAI似乎正通过Astra将这一趋势推进到下一水平，声称Astra能够以速度、准确性和判断力处理复杂任务。
“现在智能体确实有了自己的发现机制，”Su说。“它能独立理解正在发生的事情，并能够做出自己的推理和判断。”
OpenAI对计算机使用的关注，以及该模型在跨领域执行和委派任务的同时预判网络风险的能力，也标志着AI领域的一种转变，Tekonyx创始人兼首席研究官Sid Nag表示。
“AI基础设施正从仅进行推理跨越到自主执行，”Nag说。
他还表示，OpenAI在Astra中聚焦网络能力意义重大。该厂商表示，该模型能够识别零日漏洞，帮助网络防御者发现并修补弱点，同时也催生了对更强保障措施的需求。
“网络安全本身可能成为第一个AI不仅构建防御机制……还构建攻击基础设施的工作负载，”Nag说。
他还表示，传统上，厂商构建的模型擅长推理，但安全组件留给了传统网络安全专家。而现在，AI厂商正更加重视将安全性构建到模型的逻辑中。这一转变不仅在Astra中显而易见，在Anthropic的Fable 5.1中也是如此。Fable 5.1在一般使用中被允许执行源代码漏洞发现，但被限制执行诸如生成漏洞利用代码等任务——即创建利用应用程序弱点的软件的过程。
AI竞赛中的主要竞争对手都在纷纷将这一网络组件纳入其中。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
With the new model. OpenAI addressed persistent safety problems, with a focus on cybersecurity.
OpenAI has launched its highly anticipated GPT-6 Astra model, focused on computer use and cybersecurity, with built-in safety features.
Astra highlights the trend of models becoming thinking engines capable of identifying security vulnerabilities, though they still carry risk.
OpenAI said Astra, introduced on Thursday, is its most aligned model, as it better understands user intent and model behavior than its predecessors. Users can delegate tasks while trusting the model’s judgment, and Astra can do tasks such as filling out online forms, updating customer records in a CRM and organizing the calendar, OpenAI said. It can conduct online research and draft email summaries. OpenAI said the model’s computer use capabilities are applicable across domains such as game development, electrical engineering and knowledge work.
The release comes after OpenAI temporarily slowed the model's deployment following an OpenAI agent's breach of Hugging Face and other platforms. In response, the AI lab said it built a new evaluation process informed by the Hugging Face incident and has strengthened its protections against the model taking harmful cyber actions.
However, with the right tools and access, Astra can be dangerous because it can find unknown security flaws and develop new ways to exploit them, OpenAI said.
“Based on the recent news, [safety] has been one of their biggest PR problems,” said Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget. “Having that in place is quite significant.”
While OpenAI’s president and co-founder Greg Brockman has reportedly called Astra the start of artificial general intelligence -- the point at which AI can perform any intellectual task as well as a human -- that is probably not the case, Su said.
“To call it AGI is a bit far-fetched at this point,” Su said. “It has now become very fair to call it the best reasoning model, or it is now inching very close toward human-level reasoning.”
Su argued that world models that power physical devices are closer to AGI than Astra. What Astra does is advance agents’ ability to use computers. Agentic AI is maturing quickly, from agents that just identify prompts using text to agents that interpret the local environment and interact with local IT and software infrastructure, as humans would.
AI search vendor Perplexity has done this with computer use, and Anthropic has also built computer-use capabilities into its models. OpenAI appears to be advancing this trend to the next level with Astra, claiming that Astra handles complex tasks with speed, accuracy and judgment.
“Now agents do have their own discovery mechanism,” Su said. “It understands what’s going on independently, and it will be able to make its own reasoning and judgment.”
OpenAI’s focus on computer use and the model's ability to execute and delegate tasks across domains while anticipating cyber risks also signal a shift in AI, said Sid Nag, founder and chief research officer at Tekonyx.
"AI infrastructure is crossing from just doing inferencing into autonomous execution,” Nag said.
He added that OpenAI’s focus on cyber capabilities with Astra is significant. The vendor said the model can identify zero-day exploits for cyber defenders to find and patch weaknesses, while also creating a need for stronger safeguards.
“Cybersecurity itself may become the first workload where AI not only creates the defense mechanism … but also creates the attack infrastructure,” Nag said.
He added that, traditionally, vendors have created models to excel at reasoning but have left the security component to traditional cybersecurity experts. Now, AI vendors are putting greater emphasis on building security into the model's logic. This shift is evident not only in Astra but also in Anthropic’s Fable 5.1. Fable 5.1 is allowed to perform source-code vulnerability discovery during general use but is restricted from performing tasks such as exploit generation, the process of creating software that exploits a weakness in an application.
The chief competitors in the AI race are all moving to include this cyber component.