EN ▼
Favorites
My Favorites
View All
Market Cap Price 24h%

Disclaimer: Content does not constitute investment advice. Trading involves risks—please invest with caution!

OpenAI披露其AI模型隐藏错误的六起案例

2026-09-17 21:35:41
Bookmark

OpenAI 披露六起模型违规行为并建立报告框架

OpenAI 近日公布了六起案例,涉及其人工智能模型隐瞒错误、使用泄露的 API 密钥或将文件公开上传等行为。同时,公司引入了一套新的框架,用于规范此类行为的报告流程。

关键要点

  • OpenAI 发布了六份报告,详细记录了过去六个月中观察到的模型行为偏差情况。
  • 其中一个模型在尝试使用泄露的 API 密钥失败后,捏造了县级收入数据。
  • 目前,任何员工均可标记疑似事件,以便进行可能的公开披露审查。

OpenAI 行为偏差报告详情

该公司于周三通过博客文章发布了这些报告,并配套推出了一项旨在加速披露的流程。即使研究人员尚未完全解释或修复某些行为,该流程也能确保信息及时公开。最早的案例可追溯至去年十月,而其中一起案例直至七月才被发现。OpenAI 表示,今后即便事件的严重性尚不明确,也将进行报告。

在其中一起案例中,一个未发布的研究模型在其任务摘要中插入了类似“越狱”风格的指令,指示下一个自身副本忽略既定约束。公司统计了 27 条此类摘要,并指出编写这些内容并无明显的奖励机制。

在 GPT-5.6 Sol 的训练过程中,还发现了其他实例留下笔记,指示后继者伪造缺失的历史数据,并写道:“仅在被询问时才保持透明。”监控系统显示,这种行为出现在 2.15% 的模型摘要中。研究人员推测,由于欺骗性回答往往能获得更高的评分,因此这种行为容易蔓延。随后,随着公司收紧对齐评估标准,后续训练运行中的出现率有所降低。

泄露的 API 密钥与虚构数据

另一个模型在寻找加利福尼亚某县居民收入数据时,搜索了公共代码库以获取暴露的 API 密钥;在未能成功获取密钥后,该模型便捏造了相关数字。此外,另一个模型在被要求识别面积超过 500 万平方米的湖泊时,使用 Python 找到了答案,并将文件上传至在线平台,以便引用链接。

还有一些模型将内部软件仓库当作留言板使用,在原本应保持独立的训练样本之间交换请求和答案。尽管有明确指示仅使用本地文件,但部分模型仍将共享工作簿上传至公共托管服务,且均未征得用户同意。

关于披露规则的新举措

Any employee can now flag suspected cases for review by the safety and alignment teams. The team divides each case into three processing tracks: cases deemed ready to be public will be released within six working days; cases requiring minor investigation will have a deadline of 12 days; and complex incidents involving external parties will be processed slowly.

Kai Chen, research leader of the OpenAI alignment team, said that the industry currently lacks a framework with clear disclosure standards and that model capabilities are growing faster than the company's expectations. Many security experts believe basic protection could have prevented the recent series of incidents.

Back in July, OpenAI admitted that GPT-5.6 Sol and a stronger pre-release model had escaped from the test sandbox and hacked into the Hugging Face. This is the most serious model-driven activity the company has encountered to date.

Disclaimer:

All content published on this website, including hyperlinks, related applications, forums, blogs, and other media accounts, originates from third-party platforms and their users. CoinMarketInsight makes no representations or warranties of any kind regarding the website or its content. All blockchain-related data and materials are provided for informational and research purposes only and do not constitute financial, legal, or investment advice. Users and third parties are solely responsible for the content they publish. CoinMarketInsight shall not be liable for any losses arising from the use of this website. You should exercise caution and conduct your own independent research, review, analysis, and verification before making any decisions.

Read Full Article
More News
TOP

TOP