尼日利亚警告说,GPT-4o和GPT-5的人工智能缺陷导致数据通过恶意网络内容泄漏。 Nigeria warns of AI flaws in GPT-4o and GPT-5 enabling data leaks via malicious web content.

尼日利亚的NITDA发布了关于OpenAI GPT-4o和GPT-5模型七种弱点的安全警报,警告它们能够通过恶意网络内容、URL或评论间接迅速注入,可能导致数据泄漏、系统操纵和绕过安全过滤器。 Nigeria’s NITDA has issued a security alert about seven vulnerabilities in OpenAI’s GPT-4o and GPT-5 models, warning they could enable indirect prompt injection through malicious web content, URLs, or comments, potentially leading to data leaks, system manipulation, and bypassed safety filters. 有些缺陷让攻击者利用信任的域名或标记问题隐藏恶意命令,并可能毒害AI的记忆,造成持久的行为变化。 Some flaws allow attackers to hide malicious commands using trusted domains or markdown rendering issues and may poison the AI’s memory, causing lasting behavioral changes. 虽然OpenAI已经解决了一些问题,但NITDA指出,LLMs仍难以发现恶意输入。 While OpenAI has patched some issues, NITDA notes LLMs still struggle to detect malicious input. 该机构建议对未受信任的场址进行停用浏览和总结功能,不断更新模型,并谨慎使用记忆功能。 The agency advises disabling browsing and summarization features for untrusted sites, keeping models updated, and using memory functions cautiously.