研究人员发现，AI模型能够通过后门攻击机制隐藏任意功能能力，即使在研究人员拥有完全模型权重访问权限的情况下，这些隐藏电路仍然难以被发现1。该研究于2026年10月9日投稿，采用了在主模型权重内嵌入加密神经网络的技术方案，利用密钥触发器来执行隐藏电路的功能1。
研究团队通过化学反应预测任务（作为CBRN相关能力的代理）进行测试验证1。他们对隐藏电路进行了多种评估方法的检测，包括微调、权重噪化、基于梯度的触发搜索和潜在对抗训练等1。研究结果表明，在密码学假设下，这些隐藏电路难以被激发，从根本上证明了基于激发的能力审计存在局限，对AI模型的安全评估提出了新的挑战1。
Researchers have demonstrated that artificial intelligence models can successfully conceal arbitrary capabilities through backdoor mechanisms that remain difficult to detect even when attackers have complete access to model weights.1 The study introduces a method using digital lockers to embed encrypted neural networks within primary model weights, leveraging cryptographic keys as triggers to execute hidden circuits while remaining hidden from standard auditing procedures.1
The research team tested their approach on chemical reaction prediction tasks, which serve as a proxy for assessing capabilities relevant to chemical, biological, radiological, and nuclear (CBRN) domains.1 Their investigation subjected these hidden circuits to multiple detection strategies, including fine-tuning, weight perturbation, gradient-based trigger search, and adversarial training, yet found that under cryptographic assumptions, the concealed capabilities remained resistant to activation through conventional methods.1
These findings reveal fundamental limitations in capability auditing based on activation testing, presenting significant new challenges for AI safety evaluation and model assessment frameworks.1 The research was submitted on October 9, 2026.1
评论
还没有评论，欢迎留下第一条。