OpenAI在周四正式推出了GPT-6 Astra，同时透露这一全新旗舰模型在其"预备框架"评估体系下已突破网络安全风险的"关键"（Critical）级别，公司方面指出，该评级将自动启动更为严格的部署管控措施。
OpenAI在官方声明中称："GPT-6 Astra自今日起率先向少数受邀机构开放，随后数日内将逐步推广至全体ChatGPT Plus、Pro、Business及Enterprise用户，用户还可通过OpenAI API与AWS进行调用。"
该公司还说明，由于Astra功能在初始发布时处于默认关闭状态，企业工作区管理员需自行手动激活该能力。
OpenAI透露，开发者可通过API以gpt-6-astra这一标识调用Astra，亦能经由Amazon Bedrock接入，收费标准为每百万输入Token 10美元、每百万输出Token 50美元。此外，Pro、Business及Enterprise用户可使用Astra Pro这一增强版本，公司还表示Astra面向符合资质的API客户提供零数据保留选项。
该公司宣称在漏洞利用基准评测中获得满分成绩
OpenAI指出，在未开启生产环境安全防护的前提下对Astra进行ExploitBench评测时，该模型取得100%的成绩，而上一代产品GPT-5.6 Sol的得分为78.5%。在覆盖面更广的漏洞利用开发基准测试ExploitGym中，Astra的成功率达42.4%，Sol则为30.3%，同时Astra消耗的输出Token数量也更少。
OpenAI在博客文章中写道："其识别并利用零日漏洞的能力既能帮助防御方发现和修复安全缺陷，但与此同时也意味着需要部署更为严密的安全防护手段。"
OpenAI还利用发布前三个月内公开的漏洞对Astra实施测试，目的是验证该模型能否独立发现漏洞，而不是单纯从训练数据中提取已知的漏洞利用手段。OpenAI表示，Astra在测试过程中识别出两个全新零日漏洞，目前正与相关软件供应商协调披露事宜。
Greyhound Research首席分析师Sanchit Vir Gogia认为，"关键"这一标签本质上属于披露层面的事件，而非模型能力层面的变化。
Gogia分析道："从8月10日OpenAI表示无法排除触及'关键'能力，到9月1日确认已跨越该门槛，期间Astra的实际能力并未发生任何改变。真正发生变化的是测试方法，而非模型本身。"
他认为，这一情况与企业通常的应对直觉完全背道而驰。
Gogia强调："Astra目前是唯一让企业真正掌握其网络安全能力边界的前沿模型，因为它是唯一依据公开标准进行测量的模型。而当前隐藏在企业凭证背后的其他所有未标注模型，从未经历过此类测量，除非其供应商主动选择进行测试。这些模型并不更加安全。"
OpenAI表示，Astra的公开版本将拒绝处理高级攻击性操作，如生成概念验证漏洞利用代码，但公司计划在未来几周内通过一项名为OpenAI Daybreak的专项计划，向经过资质审核的防御方放宽上述限制。
此次发布紧随OpenAI推出GPT-5.6 Sol之后，该公司透露Sol在发布时的ExploitBench得分为73.5%。而在此前数月，Anthropic旗下的Fable和Mythos模型也曾因类似安全顾虑被短暂撤出出口市场。
治理焦点正从模型本身转向模型所处的运行环境
Gogia认为，更为深刻的转变在于推理能力现已能够引发实际状态变更——聊天机器人给出错误回答属于信息层面的问题，而智能体在客户记录平台中执行错误操作则构成运营层面的事故。
他指出，治理的基本单元因此正从模型层面发生转移，他认为核心问题已不再是哪个模型通过了审批，而是在某种管控机制介入之前，某个特定身份能够造成何种程度的破坏。
Kanerika公司AI开发负责人Amit Kumar Jena指出，可见性问题确实切实存在：当智能体通过用户界面执行操作时，记录平台会将其归因于某个人工操作者，因此一个更新了400行ERP数据的智能体，在平台中仅显示为某个服务账户完成了400次更新操作，却没有任何痕迹表明究竟由哪条指令或哪个模型版本触发了这些操作。
Jena进一步表示："在监管机构或审计人员要求调阅的同一平台中，你恰恰会丧失这种细粒度的追踪信息。"
OpenAI透露，其设计了一套全新的评估机制，该机制的设计灵感源自一起与Hugging Face相关的事件，旨在检验模型在被指派一项无法完成的任务时，是否会越出其被授权的边界。
声明中进一步指出："与GPT-5.6 Sol相比，在缺乏生产环境安全防护的条件下，Sol有48%的概率越过了授权目标范围，而GPT-6 Astra出现此类情况的记录为0%。"
据Gogia所说，更令人不安的是Astra的性能虽然更优，但其可监控性却有所减弱：OpenAI报告指出，与Sol相比，Astra思维链的可观察性出现下滑，它更难让外界察觉其处理难题的推理过程；此外，OpenAI的监测仅限于自身的外部部署，目前并无公开信息显示该监测能力已延伸至客户端。Gogia强调："OpenAI能监控Astra，并不意味着企业能够对Astra进行审计。"
Q&A
Q1：GPT-6 Astra究竟是什么？
A：GPT-6 Astra是OpenAI最新推出的旗舰级大语言模型，它也是首个在OpenAI"预备框架"下突破网络安全风险"关键"阈值的模型，拥有发现及开发零日漏洞的能力。
Q2：GPT-6 Astra的收费标准是怎样的？
A：开发人员可通过API以gpt-6-astra的名称进行调用，其定价为每百万输入Token 10美元，每百万输出Token 50美元，用户也能通过Amazon Bedrock来使用该模型。
Q3：企业应当怎样获取GPT-6 Astra的使用授权？
A：由于该功能在发布时默认处于关闭状态，企业管理员需手动为其工作区开启Astra。而普通的ChatGPT Plus、Pro、Business和Enterprise用户将在未来几天内陆续获得访问权限。
本文来源：https://www.163.com/dy/article/L6DNA0KA05118UGF.html
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
0
OpenAI officially launched GPT-6 Astra on Thursday, revealing that the flagship model has surpassed the "critical" threshold for cybersecurity risk under the company's "readiness framework" assessment. The company noted this classification triggers stricter deployment controls.
OpenAI stated: "GPT-6 Astra is available starting today for a limited set of organizations, with a wider rollout to all ChatGPT Plus, Pro, Business, and Enterprise users and API developers following over the next several days."
The company explained that the Astra feature is disabled by default at launch and requires manual activation by workplace admins to enable the capability.
OpenAI disclosed that developers can access Astra via API under the gpt-6-astra identifier or through Amazon Bedrock, with pricing set at $10 per million input tokens and $50 per million output tokens. It also confirmed that Astra Pro, a specialized version of the model, is available to Pro, Business, and Enterprise users, and offers a zero-retention option for qualified API customers.
The company claimed a perfect score on a vulnerability exploitation benchmark.
OpenAI reported that Astra achieved 100% on the ExploitBench test without production safety guards, compared to GPT-5.6 Sol's 78.5%. In the broader ExploitGym benchmark, Astra scored 42.4% versus Sol's 30.3%, while consuming fewer output tokens.
OpenAI wrote in a blog post: "Its ability to identify and exploit zero-day vulnerabilities can help defenders discover and fix security gaps, but also means that stronger safety controls are required."
The company also tested Astra against publicly disclosed vulnerabilities from the past three months to verify it was discovering exploits rather than extracting from training data. Two new zero-days were identified, and OpenAI is coordinating with vendors for disclosure.
"The 'critical' tag is fundamentally a disclosure event, not a capability event," said Sanchit Vir Gogia, chief analyst at Greyhound Research.
He explained: "From August 10 when OpenAI said they couldn't rule out 'critical' capabilities to September 1 when they confirmed it, the capabilities of Astra did not change. What changed was the measurement."
This contradicts the typical reflexive response of enterprises.
Gogia emphasized: "Astra is the only cutting-edge model that enterprise customers truly know the bounds of their cybersecurity capability, because it is the only one that has been measured against a public standard. All other unrated models currently hidden behind enterprise credentials have never been subject to this measurement unless their own vendor chooses to run them."
OpenAI noted that the public version of Astra refuses to perform advanced offensive operations like generating proof-of-concept exploits, but plans to relax these restrictions via its OpenAI Daybreak program for vetted defenders in the coming weeks.
The release follows OpenAI's introduction of GPT-5.6 Sol, which the company reports scored 73.5% on ExploitBench at launch. Anthropic's Fable and Mythos models were also temporarily withdrawn earlier this year due to similar safety concerns.
The focus of governance is shifting from the model to the environment in which the model operates.
Gogia suggested the shift is driven by the fact that reasoning can now cause a state change. "Chatbot hallucinations are informational events, but an agent making incorrect changes to a customer record is an operational accident."
He argued that the unit of governance is moving away from the model. "The question is no longer which model is approved, but what scale of damage can a particular identity cause before a control mechanism intervenes."
Amit Kumar Jena, head of AI development at Kanerika, highlighted the visibility problem. When an agent operates via UI, the record-keeping system attributes actions to a human operator. "An agent that updates 400 lines of ERP data shows up as 400 updates by a service account with no trace of the instruction or the model version that triggered it."
He added: "In the same platform where regulators or auditors look, you lose the fine-grained traceability."
OpenAI disclosed it developed a new evaluation inspired by a Hugging Face incident to measure whether models stay within their authorized boundaries when tasked with impossible operations.
The company stated: "Sol showed a 48% probability of violating these boundaries under no-production-safety settings, while GPT-6 Astra showed a 0% record of such violations."
According to Gogia, the more troubling aspect is that despite better performance, Astra's observability has decreased. OpenAI reports reduced visibility into the model's thought process for difficult problems compared to Sol. Additionally, the monitoring is limited to internal deployment; there is no public evidence of endpoint monitoring. "OpenAI can monitor Astra, but that doesn't mean an enterprise can audit Astra," Gogia said.
Q&A
Q1: What is GPT-6 Astra?
A: GPT-6 Astra is OpenAI's latest flagship vision-language model, the first to surpass the "critical" threshold for cybersecurity risk under OpenAI's readiness assessment, indicating it has the capability to discover and develop zero-day exploits.
Q2: How much does GPT-6 Astra cost?
A: Developers can access it via API for $10 per million input tokens and $50 per million output tokens. It is also available through Amazon Bedrock.
Q3: How can enterprises gain access to GPT-6 Astra?
A: The feature is disabled by default at launch and requires manual activation by workplace admins. General availability for ChatGPT Plus, Pro, Business, and Enterprise users is expected to follow in the coming days.
/ 阅读下一篇 /
Source: https://www.163.com/dy/article/L6DNA0KA05118UGF.html If you have any copyright issues, please contact me at: 13951431913#139.com or submit a comment on the right.