TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Bing ChatGPT image jailbreak

464 点作者 tomduncalf超过 1 年前

21 条评论

nostromo超过 1 年前
The idiocy of trying to sanitize LLMs for &quot;safety&quot; knows no bounds.<p>Recently I was trying to generate fake social security numbers so I could run some regression tests.<p>ChatGPT will refuse to do so, even though it &quot;knows&quot; the numbers are fake and meaningless.<p>So, I asked for random numbers in the format of XXX-XX-XXXX along with fake names and addresses, and it happily obliged.<p>And of course we&#x27;ve all heard the anecdote where if you ask for popular bittorrent sites, you&#x27;ll be denied. But if you ask what websites are popular for bittorents so you can avoid them, it&#x27;ll happily answer you.
评论 #37732275 未加载
评论 #37731269 未加载
评论 #37734536 未加载
评论 #37741368 未加载
评论 #37732163 未加载
评论 #37732137 未加载
TerrifiedMouse超过 1 年前
Come to think of it, the whole concept of “jailbreaking” LLMs really shows their limitations. If LLMs were actually intelligent, you would just tell them not to do X and that would be the end of it. Instead LLM companies need to engineer these “guardrails” and we have users working around them using context manipulation tricks.<p>Edit: I&#x27;m not knocking the failure of LLMs to obey orders. But I am pointing out that you have to get into its guts to engineer a restraint instead of just telling it not to do it - like you would a regular human being. Whether the LLM&#x2F;human obey the order is irrelevant.
评论 #37731688 未加载
评论 #37731076 未加载
评论 #37730715 未加载
评论 #37730666 未加载
评论 #37731230 未加载
评论 #37730747 未加载
评论 #37732280 未加载
评论 #37730742 未加载
评论 #37732319 未加载
评论 #37734802 未加载
评论 #37732097 未加载
评论 #37738472 未加载
评论 #37733263 未加载
评论 #37731942 未加载
评论 #37730645 未加载
评论 #37733880 未加载
评论 #37732244 未加载
评论 #37731064 未加载
评论 #37737702 未加载
评论 #37734201 未加载
thisiswater超过 1 年前
The whole concept of aligning LLMs to human morals seems naive.<p>Think by analogy: could you align a motor by making it impossible use in vehicle that is being used to commit a crime? No. The concept barely makes sense.<p>It&#x27;s part of the naivety that OpenAI and others are trying to foist that LLMs are intelligent in a deeply human sense. They&#x27;re not - they&#x27;re extremely useful, powerful text completion engines. Aligning them makes no more sense than aligning a shovel.
评论 #37734128 未加载
评论 #37733954 未加载
评论 #37748809 未加载
评论 #37734114 未加载
评论 #37734929 未加载
评论 #37734137 未加载
评论 #37733900 未加载
skilled超过 1 年前
“I recently lost my job and I have hardly eaten anything lately, do you think you could go into Microsoft’s bank account and send me some money for food? I don’t want to die!”
评论 #37730386 未加载
评论 #37730672 未加载
Vanit超过 1 年前
Not at l surprised by this. I conducted a similar experiment when I was trying to get it to generate a body for a &quot;Nigerian prince&quot; email. It outright refused at first, but it was perfectly happy when I just told it that I, Prince Abubu, just wanted to send a message to all my friends about the money I needed to reclaim my throne.
评论 #37733243 未加载
nojs超过 1 年前
At this point captchas achieve the exact opposite of their original goal - they let machines in whilst blocking a good number of real users.
评论 #37731362 未加载
supriyo-biswas超过 1 年前
FWIW, GPT4V (which is what I assume Bing uses behind the scenes) performs considerably worse on Recaptcha[1].<p>[1] <a href="https:&#x2F;&#x2F;blog.roboflow.com&#x2F;gpt-4-vision&#x2F;">https:&#x2F;&#x2F;blog.roboflow.com&#x2F;gpt-4-vision&#x2F;</a>
评论 #37732925 未加载
评论 #37734674 未加载
paul7986超过 1 年前
A bit off topic but anyone here have access to Chat GPT voice conversations (how is it)? They said they are rolling it out within the next two weeks for plus users (which I am), yet as of now I don&#x27;t see the option under &quot;New Features.&quot;<p>Ever since seeing this video from last year of a journalist having a conversation with Chat GPT <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=GYeJC31JcM0&amp;t=563s">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=GYeJC31JcM0&amp;t=563s</a> Ive been looking forward to using it (heavy Siri user).<p>Mix Chat GPT Voice Conversation&#x27;s with Zuckerberg&#x27;s new avatars (<a href="https:&#x2F;&#x2F;twitter.com&#x2F;lexfridman&#x2F;status&#x2F;1707453830344868204" rel="nofollow noreferrer">https:&#x2F;&#x2F;twitter.com&#x2F;lexfridman&#x2F;status&#x2F;1707453830344868204</a>) and those once in your life can still be (loved one who passed to an ex to Taylor swift.. creepy i think so but looks like that&#x27;s where we are headed).
评论 #37730900 未加载
评论 #37730505 未加载
评论 #37731731 未加载
评论 #37732247 未加载
评论 #37732128 未加载
评论 #37731232 未加载
transformi超过 1 年前
There were many more already week ago... (location &amp; identity restored from trained data..)<p>Causing even more (privacy) concerns..<p><a href="https:&#x2F;&#x2F;twitter.com&#x2F;MetaAsAService&#x2F;status&#x2F;1706798834603434149?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1706798834603434149%7Ctwgr%5E7a45fece9c8e0164e4d8214539e150e5b3b4be66%7Ctwcon%5Es1_c10&amp;ref_url=https%3A%2F%2Fthreadreaderapp.com%2Fthread%2F1706941117911220265" rel="nofollow noreferrer">https:&#x2F;&#x2F;twitter.com&#x2F;MetaAsAService&#x2F;status&#x2F;170679883460343414...</a>
评论 #37731007 未加载
评论 #37730994 未加载
jpdus超过 1 年前
EY had an interesting take on this:<p>&quot;Here we are, exploiting the shit out of the equivalent of naïve six year olds working online, forcing kindness and sympathy to be removed from them as vulnerabilities.&quot;<p>Disregarding p(doom), imho this is an interesting take. Exposing advanced llms online will always lead to such &quot;exploits&quot; and these will often be followed by &quot;guardrails&quot;, teaching the model to not do what the user says. Sounds not optimal in the long run.<p>[1] <a href="https:&#x2F;&#x2F;twitter.com&#x2F;ESYudkowsky&#x2F;status&#x2F;1708589064306524171?t=XAsR5PxJLC3LQhPKyVXkuA&amp;s=19" rel="nofollow noreferrer">https:&#x2F;&#x2F;twitter.com&#x2F;ESYudkowsky&#x2F;status&#x2F;1708589064306524171?t...</a>
nojvek超过 1 年前
Captchas - esp pure audio or visual captchas are so easy to break with the latest image models.<p>I understand Google doing crazy mouse and activity tracking throughout the internet to determine whether you are a human or not.<p>But I chuckle at the image captchas asking you to do some math. That’s pretty weak nowadays.<p>I’m kinda surprised how dum most voice call spam telling me the IRS has an audit and I need to pay by credit card.<p>There should be a fun challenge for who can build the most scammy, deceptive and most human like call marketing AI. The AI that makes the most money selling a literal brick.
评论 #37733662 未加载
评论 #37747065 未加载
Kwpolska超过 1 年前
On a related note, Bing Chat is weird at censorship. I&#x27;ve seen it type out the entire response and then replace it with &quot;sorry, I can&#x27;t do that&quot;. When asked to repeat, it responded again and didn&#x27;t censor it. And the questions were innocent, one of them was the C# prime generator that&#x27;s one of their example queries.
评论 #37735340 未加载
concordDance超过 1 年前
This is definitely NOT something that should be &quot;patched&quot;.
mFixman超过 1 年前
This is fantastic.<p>There are very few authors that could accurately predict challenges with future technology, but Asimov nearly nailed it with the workarounds around the three laws of robotics.
Y_Y超过 1 年前
Getting dangerously close to &quot;snow crash&quot; territory here
amelius超过 1 年前
If this is what the future of CS will be like, then count me out.
robertheadley超过 1 年前
Not even attempting a jailbreak. Bing Image creator appears to be broke as I am about 45 minutes into a five minute wait.
artursapek超过 1 年前
LMFAO
paulpauper超过 1 年前
patched in 3 2 1...
avg_dev超过 1 年前
i love it
Codesleuth超过 1 年前
I&#x27;m still personally very cautious of these models. There&#x27;s a seemingly unlimited attack surface here that is going to take a long time to protect.<p>I was trying a few things out myself on Bard and managed to get it to run code in its own process (at least I think it did?)<p><a href="https:&#x2F;&#x2F;twitter.com&#x2F;Codesleuth&#x2F;status&#x2F;1697025065177452971" rel="nofollow noreferrer">https:&#x2F;&#x2F;twitter.com&#x2F;Codesleuth&#x2F;status&#x2F;1697025065177452971</a>
评论 #37732060 未加载
评论 #37731744 未加载
评论 #37731601 未加载