OpenAI近期发布了一份报告,详细阐述其应对“对抗性蒸馏”的安全措施 1。该策略旨在防止攻击者利用前沿模型的输出训练其他模型以绕过安全防护 1。OpenAI将“对抗性蒸馏”定义为未经授权地利用前沿模型输出来训练具有相似能力但缺乏相应防护措施的新模型 1。
在技术防御层面,OpenAI通过强化GPT-6.2 Astra和Sol版本的防御机制进行了内部自适应评估 1。结果显示,针对受保护推理(即思维链或chain-of-thought)的提取攻击平均成功率从GPT-6 Astra原有的54.0%降至0.2% 1。OpenAI强调保护“受保护推理”至关重要,因为其中包含最终答案未显示的中间步骤,是极具价值的训练数据 1。
除了模型层面的防护升级,OpenAI还讨论了该技术带来的双重用途风险以及跨层级的检测与执法策略 1。自2023年夏季以来,OpenAI每周针对此类活动的执法行动涉及数千至数万个账户 1。该公司提到今年2月向国会简报了DeepSeek试图蒸馏美国前沿模型的活动,并在9月报告了与Moonshot AI相关的提取活动 1。此外,OpenAI表示支持《芯片安全法案》、《AI OVERWATCH法案》及《MATCH法案》,旨在限制对手获取硬件并保护AI能力 1。
OpenAI has released a report detailing security measures against "adversarial distillation," defined as the unauthorized use of frontier model outputs to train new models with similar capabilities but lacking corresponding safety protections 1. The company emphasizes safeguarding "protected reasoning" (chain-of-thought), noting that these intermediate steps contain valuable training data not visible in final answers 1. Since summer 2023, OpenAI’s weekly enforcement actions against such activities have involved thousands to tens of thousands of accounts 1.
Internal adaptive assessments show that strengthened defenses in GPT-6.2 Astra and Sol versions have significantly reduced vulnerabilities 1. The average attack success rate for extracting protected reasoning dropped from 54.0% in GPT-6 Astra to below 0.2% in the updated GPT-6.2 Astra model 1. The report also discusses dual-use risks posed by this technique in cybersecurity and biological fields, and outlines multi-layered detection and enforcement strategies 1.
On the policy front, OpenAI stated it briefed Congress in February regarding DeepSeek’s attempts to distill American frontier models, and reported extraction activities linked to Moonshot AI in September 1. The company supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act to restrict adversaries’ access to hardware and protect AI capabilities 1.
评论
还没有评论,欢迎留下第一条。