GPT-2 模型复现 (AML.CS0007)

一句话通俗理解

OpenAI 不敢放出的 GPT-2 完整模型，被社区用「猜词游戏」一点点复现了出来。

案例速览

事件经过

OpenAI 构建了 GPT-2 语言模型，因担心被滥用（冒充他人、生成虚假信息）而分阶段发布。社区研究者利用逐词查询预测的方式，成功复现了完整模型的行为——证明对公开 API 的黑盒访问足以复现被「雪藏」的模型。

运用的 ATLAS 技术

以下攻击链映射来自 MITRE ATLAS 官方数据（AML.CS0007 的 employs 关系），按官方步骤编号排序；技术名称链接至本库对应技术文档，引用文字为官方对本次运用的描述。

S00

• 搜索开放技术数据库 (AML.T0000)

  Using the public documentation about GPT-2, the researchers gathered information about the dataset, model architecture, and training hyper-parameters.

S01

• 模型 (AML.T0002.001)

  The researchers obtained a reference implementation of a similar publicly available model called Grover.

S02

• 数据集 (AML.T0002.000)

  The researchers were able to manually recreate the dataset used in the original GPT-2 paper using the gathered documentation.

S03

• AI 开发工作区 (AML.T0008.000)

  The researchers were able to use TensorFlow Research Cloud via their academic credentials.

S04

• 收集 AI 产物训练代理 (AML.T0005.000)

  The researchers modified Grover’s objective function to reflect GPT-2’s objective function and then trained on the dataset they curated using used Grover’s initial hyperparameters. The resulting model functionally replicates GPT-2, obtaining similar performance on most datasets. A bad actor who followed the same procedure as the researchers could then use the replicated GPT-2 model for malicious purposes.

关联战术

• 侦察 (AML.TA0002)
• 资源开发 (AML.TA0003)
• AI 攻击适配 (AML.TA0001)

参考资料

以下为 MITRE ATLAS 官方案例收录的参考来源：

• Wired Article, “OpenAI Said Its Code Was Risky. Two Grads Re-Created It Anyway”

• Medium BlogPost, “OpenGPT-2: We Replicated GPT-2 Because You Can Too”

• MITRE ATLAS 官方案例页 (AML.CS0007)

版本: ATLAS 2026.09 | 更新: 2026-10-24 | 模板: ATLAS 案例研究 v1.0