ZDNET的关键外卖
- 一个实验工具从聊天中找到删除的信息。
这对网络安全调查很有帮助。
- 也表明当地人工智能安全存在局限性。
工具如何运作
许多有隐私意识的人工智能用户更喜欢在设备上本地运行助手，因为他们的数据不会被发送到云中，在云中可能会被偷窥眼睛或数据被破坏。 相反，聊天会话使用您的设备内存。
即使您自己删除了聊天内容。 Georgia Tech研究者创建了ORISA，这个工具可以搜索那些内存以获取线索来删除对话。 他们为自己的实验设计了假记名AI助手，使用ChatterUI Android应用，给了它一套100个关于一个项目的虚构信息，然后删除了这些信息。
更多来自 ZDNET
还：您可以在您的 Android 或 iOS 设备上运行本地AI
研究人员使用ORISA，然后搜索了设备内存，以查找该信息的后遗症，这些后遗症可以作为上下文来重启会话。 ORISA找到了足够的可用信息，以创建新的工作会话，从最初的100集中得出了26个虚构的事实。 研究者解释道，这比最初的14次助教所访问的更多
“这可以帮助确定哪些人与AI助理分享信息，以及他们得到了哪些援助”，博士生Ken Xu和研究的主要作者称是实验的作者。
对你意味着什么
ORISA是用来填补网络安全漏洞的。 研究者指出，当地人工智能的一个缺点是缺乏“集中的伐木和执法”。 这说明坏角色可以使用微调犯罪模型
迄今为止，这一现实使调查人员难以使用法医记忆技术来取回这一活动的有用证据。
还有：勒诺沃的RTX Spark笔记本电脑通过了我完全疯狂的AI编码测试
研究者写道，“ORISA利用批量符号和在进程内存中保存的密钥值缓存来重建符号序列并恢复会话级别注意状态”。
这意味着这样的工具与那些试图阻止网络犯罪的人最为相关，而不是与那些挖掘你个人数据的人最为相关。 称，如果黑客重现了这样的技术或者他们掌握了像ORISA本身这样的工具，他们理论上可以使用它从您的聊天中收集被删除的信息，并利用这些信息作为上下文重新创建您的查询。 但ORISA的功效也取决于时间 — — 研究者指出，随着时间的流逝，恢复被删除的信息更加困难，因为手机的记忆被更新的活动所覆盖。
无论如何，我们的长期建议仍然是：不要与聊天员分享敏感信息
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
ZDNET’s key takeaways
- An experimental tool recovered deleted information from chats.
- It’s helpful for cybersecurity investigations.
- It also demonstrates limitations in perceived local AI security.
How the tool works
Many privacy-conscious AI users prefer to run assistants locally on their devices, because their data isn’t sent out to the cloud, where it could be vulnerable to prying eyes or data breaches. Instead, the chat session uses your device memory.
But, as it turns out, some information from those sessions can stick around in your phone’s working memory even if you’ve deleted the chats themselves. Georgia Tech researchers created ORISA, a tool that searches that memory for clues to deleted conversations. They designed a fake note-keeping AI assistant for their experiment using the ChatterUI Android app, gave it a set of 100 fictional pieces of information about a project, and then deleted that information.
More from ZDNET
Also: You can run local AI on your Android or iOS device
Using ORISA, the researchers then searched the device memory for remnants of that information that could be used as context to restart the session. ORISA found enough usable information to create a new working session with 26 fictional facts from the original set of 100. That’s more than the 14 the original assistant session had access to, “because the rest had fallen outside its active memory,” the researchers explained.
“It can help establish what information someone shared with an AI assistant and what assistance they received,” Ph.D. student Ken Xu, and lead author of the study, said of the experiment.
What it means for you
ORISA was developed to fill a cybersecurity gap. The researchers noted that a downside of local AI is its lack of “centralized logging and enforcement.” That means bad actors can fine-tune models for crimes, for example, and then make the evidence untraceable by deleting them after they’ve used them.
Until now, that reality has made it difficult for investigators using forensic memory techniques to recover useful evidence of that activity.
Also: Lenovo’s RTX Spark laptop passed my completely bonkers AI coding test
“ORISA leverages the batch tokens and key-value (KV) caches preserved in process memory to reconstruct token sequences and restore session-level attention states,” the researchers wrote.
That means a tool like this is most relevant to those trying to stop cybercrime rather than to those digging into your personal data. That said, if hackers recreate a technology like this or get their hands on a tool like ORISA itself, they could theoretically use it to gather deleted information from your chats and recreate your queries using that information as context. But ORISA’s efficacy also depends on time — the researchers noted that it’s harder to recover deleted information as time goes on, because your phone’s memory gets overwritten by newer activity.
Either way, our long-standing advice remains: Don’t share sensitive information with chatbots, even in a local session.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.