在人工智能领域，安全问题一直是开发者和用户关注的焦点。近日，人工智能专家Andrej Karpathy在社交媒体上发表了一篇关于大型语言模型（LLM）安全的文章，揭示了一个潜在的安全漏洞，即特殊token的注入攻击。Karpathy指出，大型语言模型中的分词器可能会将用户输入中的特殊token解释为实际的特殊token，这可能会导致模型无法正确处理输入，甚至可能被用来进行未授权的数据访问或修改。
Karpathy强调，为了提高大型语言模型的安全性，开发者应该避免在编码和解码过程中通过解析字符串的方式来处理特殊token，而是应该通过编程方式显式地添加这些token。他还提到，虽然这是一个微妙且记录不全的问题，但预计有50%的代码都存在这种问题导致的bug。
对此，Google DeepMind的科学家Lucas Beyer表示，他们在新工作中已经提升了安全机制，并指出尽管这可能会带来一些麻烦，特别是在支持多个tokenizer时，但这是确保模型安全性的必要步骤。
专家们一致认为，最小特权原则是提高人工智能系统安全性的关键。这意味着在设计系统时，应该限制其功能，只允许进行必要的操作，以减少意外后果的可能性。
大型语言模型的安全性是一个不断发展的领域，随着技术的进步，开发者需要不断更新他们的安全措施，以确保这些模型能够安全可靠地被使用。
英语如下：
News Title: “AI Industry Leader Reveals Security Flaws in Large Language Models: Easy SQL Injection Attacks”
Keywords: Security Vulnerabilities, AI Models, Andrej Karpathy
News Content: In the realm of artificial intelligence, security issues have always been a focal point for developers and users. Recently, AI expert Andrej Karpathy published an article on social media about the security of large language models (LLMs), revealing a potential security vulnerability known as special token injection attacks. Karpathy pointed out that the tokenizer within large language models may interpret special tokens within user inputs as actual special tokens, which could lead to the model misinterpreting the input and potentially being exploited for unauthorized data access or modification.
Karpathy stressed the importance of developers avoiding the parsing of special tokens through string-based methods during encoding and decoding processes, instead opting for explicit programming to add these tokens. He also mentioned that, while this is a subtle and under-documented issue, it is expected that 50% of the code could be affected by bugs stemming from this problem.
In response, Google DeepMind scientist Lucas Beyer stated that their new work has enhanced the security mechanisms, noting that while this may cause some inconvenience, especially when supporting multiple tokenizers, it is a necessary step to ensure the model’s security.
Experts agree that the principle of least privilege is critical in enhancing the security of artificial intelligence systems. This principle suggests that system design should restrict functionality, allowing only necessary operations to minimize the likelihood of unintended consequences.
The security of large language models is an evolving field, and as technology progresses, developers must continuously update their security measures to ensure these models can be used safely and reliably.
【来源】https://www.jiqizhixin.com/articles/2024-08-16-2
Views: 2