**Meta的Llama 2大模型安全性遭质疑，DeepKeep评估报告揭示严重问题**
据IT之家报道，知名AI安全公司DeepKeep近日发布了一份评估报告，直指Meta公司的Llama 2大语言模型在安全性方面存在显著问题。在涵盖13个风险评估类别的严格测试中，Llama 2仅通过了4项，引发了业界对AI模型安全性的关注。
报告指出，Llama 2，特别是其拥有70亿参数的7B版本，存在严重的“幻觉”问题。这意味着该模型在生成回答时，有高达48%的几率产生虚假或误导性的内容。这一比例令人担忧，因为大语言模型通常被用于各种应用，包括自动驾驶、医疗诊断和决策支持系统，其准确性与安全性直接影响到用户和公众的福祉。
DeepKeep的评估结果揭示了Llama 2在确保信息真实性和避免误导性输出方面的不足，这可能会对依赖该模型的组织和用户造成潜在的危害。Meta公司对此尚未发表正式回应，但这一情况无疑对其AI研发策略提出了挑战，同时也为整个AI行业敲响了警钟，强调了在追求技术进步的同时，必须强化模型的安全性和可信度。
随着AI技术的广泛应用，模型的可靠性和安全性已经成为衡量其价值的关键标准。DeepKeep的报告提醒业界，对于大型语言模型的开发和应用，必须加强安全测试和风险评估，以确保公众利益得到保障。
英语如下：
**News Title:** “Serious Safety Concerns with Meta’s Llama 2 Model: DeepKeep Report Finds 48% Hallucination Rate”
**Keywords:** Meta Llama 2, safety issues, high hallucination rate
**News Content:**
According to IT Home, renowned AI security firm DeepKeep recently released an assessment report highlighting significant security concerns with Meta’s Llama 2 large language model. In rigorous tests across 13 risk assessment categories, Llama 2 passed only four, drawing attention to the broader issue of AI model safety.
The report reveals that Llama 2, particularly its 7B version with 70 billion parameters, suffers from severe “hallucination” issues. This means the model generates false or misleading content in as high as 48% of its responses. This rate is alarming given that large language models are often employed in applications such as autonomous driving, medical diagnosis, and decision support systems, where accuracy and safety directly impact user and public welfare.
DeepKeep’s assessment exposes Llama 2’s shortcomings in ensuring informational authenticity and preventing误导性的 outputs, posing potential hazards to organizations and users relying on the model. Meta has yet to issue a formal response, but this situation undoubtedly poses a challenge to its AI development strategy and serves as a wake-up call to the entire AI industry, emphasizing the need for enhanced model safety and credibility alongside technological advancements.
As AI technology becomes increasingly prevalent, model reliability and security have become crucial benchmarks of their value. DeepKeep’s report underscores the necessity for stricter safety testing and risk assessments in the development and application of large language models, ensuring public interests are protected.
【来源】https://www.ithome.com/0/762/593.htm
Views: 1