Wish Wu

AI Jailbreak,
Android Root,
Bypass FaceID,
Automatic Vulnerability Mining

[pattern-f @pattern_F_]

Yes😉

[International Cyber Digest @IntCyberDigest]

‼️ BREAKING: Anthropic has embedded hidden spyware-like code in Claude Code that covertly targets Chinese users. It then sends information regarding every user by injecting it into their prompt message.

Claude Code is sending info like timezone, proxy and possible AI Lab connections into the system prompt in ways Chinese users can't notice.

A coding agent with repo and command permissions should not silently hide routing metadata inside prompts. This is a serious breach of user trust.

[Chaofan Shou @Fried_rice]

vibe coded a fuzzing ai agent last month and let it run for a week using my $200 claude max.

it then found 21 high/critical vulnerabilities in Chrome.

[Elon Musk @elonmusk]

People giving OpenClaw root access to their entire life

[zR @zRdianjiao]

GLM-5 is now open-source under the MIT License.

Huge thanks to all the developers who’ve supported us along the way — we truly appreciate it.

[karminski-牙医 @karminski3]

就这??? 马斯克你认真的吗?

来看 Grok4 实测! 我原本打算用新试题, 突然转念一想, 万一Grok4 延续了 Grok3 的辉煌传统怎么办? 于是直接用了经典到都出包浆的我的那个20小球在七边形里面弹跳的实体快速来了一遍.

结果, 3次生成代码中, 2次 Grok4生成的代码甚至有语法错误. 唯一一次成功的是这个样子. 为了给不了解这个测试的朋友做对比, 我放了 DeepSeek-R1 作为参考, 注意哦，这个甚至不是 DeepSeek-R1-0528, 而是今年年初的那个老版本 R1....

从目测来看, 大概是今年第一季度所有大模型中, 写代码能力的中间水平, 接近GPT-4o 或 kimi-1.5-long-thinking 之间. 根本不是第一梯队的水平, 更别提跟现在的 Claude-4 或者 Gemini-2.5-pro 比了.....

建议 AIME25 那个数学能力也谨慎看待, 说不定又是一个过拟合的结果.

更多测试我随后放出, 希望 Grok4 能打我的脸. 但这编程水平....呵呵.

#Grok4

[clem 🤗 @ClementDelangue]

Our science team has started working on fully reproducing and open-sourcing R1 including training data, training scripts,...

Full power of open source AI so that everyone all over the world can take advantage of AI progress! Will help debunk some myths I’m sure too.

Thanks @deepseek_ai!

[Junxian He @junxian_he]

We replicated the DeepSeek-R1-Zero and DeepSeek-R1 training on 7B model with only 8K examples, the results are surprisingly strong.

🚀 Starting from Qwen2.5-Math-7B (base model), we perform RL on it directly. No SFT, no reward model, just 8K MATH examples for verification, the resultant model achieves (pass@1) 33.3% on AIME, 62.5% on AMC, and 77.2% on MATH, outperforming Qwen2.5-math-7B-instruct and being comparable to PRIME and rStar-MATH that use >50x more data and more complicated components.
🚀 Increased CoT length and self-reflection emerge

We share the details and our findings in the blog:
https://t.co/DoupDsB4hA
Training code and implementation details here: https://t.co/6TRqSr2eBG

[Taelin @VictorTaelin]

mini rant:

the illusion LLMs are intelligent comes from their massive scale. it is hard to visualize, but these things memorized the WHOLE internet. everything you've ever asked it, either has been solved before, or is a simple combination of existing solutions. but that's still an illusion.

when LLMs face a problem that require a NEW solution - one of original shape, one it has never seen before - they fail. it is that simple. that's what my example shows. i took as simple problem - inverting a binary tree - and added a few constraints to make sure the solution is unique and not in the dataset, forcing it to actually SOLVE it itself. and, surprise - it can't!

and I must stress this isn't about THIS problem. but about all. LLMs can't solve ANY problem, at all. it can only spit a memorized solution. if nobody posts this solution online, not even GPT-6, opus-5, or o3, will be able to solve *this very prompt*. I'm betting on that.

the inability to create new solutions imply LLMs won't invent new science. yes, they will completely change the world as we know it. they'll have a higher impact than computers and the internet. but, unless a new kind AI emerges, we're still on our own when it comes to curing cancer or making superconductors.

[Wish Wu @wish_wu]

Thank you for the invitation from @GEEKCONTOP. I hope it continues to improve and become a top-tier event in the security community.

[GEEKCON @GEEKCONTOP]

https://t.co/X16yWQ9uvT

[小島秀夫 @Kojima_Hideo]

本当は原作を読んで欲しいところ。原作ファンには、テンセント版がおすすめかも。

[Webcam Motion Capture @webcammocap]

As of today, we have now reached over 10,000 subscribers. We truly appreciate your support!

We are excited to announce that we will soon be releasing the Webcam-based Face Expression Tracking feature, as demonstrated in the video!

Stay tuned for more updates😊

[Pinjia He @PinjiaHE]

Excited about my PhD student @youliang_yuan's work. We find that chat in cipher can bypass the safety alignment techniques of LLMs and LLMs seem to have a "secret cipher" inside.
https://t.co/UnNjviG4Vf

[Maddie Stone @maddiestone]

✨🤖 New RCA by @tehjh up! CVE-2023-20963 is a 0-day in Android's Parcel serialization/deserialization which was used in-the-wild by the pinduoduo app. #itw0days

https://t.co/7Jah3SRvpu

[Emmanuel Macron @EmmanuelMacron]

谢谢你，广州。
法中友谊万岁！
Merci Canton.
Vive l’amitié entre la Chine et la France !

[zh1x1an1221 @zh1x1an1221]

weird UAF :)

[Wish Wu @wish_wu]

因为那些“施救”的国家就没真打算“施救”。

[Sir @YahuiWang9]

@DemEnthusiast @wangzhian8848 最近20年号称被“拯救”过的国家的现状都太惨了啊，一地鸡毛，根本找不出几个从专制过渡到民主过渡的好的正面案例，所以大家不信任也不足为奇。（我说的是近20年以来从专制改变成民主的，不是那些从一开始成立时候就是民主体制的国家。另外，如果有人说宁可饿死、战死、焦土化掉也值得的当我没说）

[Wish Wu @wish_wu]

我曾经是个反贼，但被一些人只从立场出发的双标、田忌赛马氏造谣、拿捕风捉影的东西做莫须有的指责、只讲情绪不讲科学理性的行为给气成了“粉红”。

[Wish Wu @wish_wu]

@VYSEa @victorchan2022 这真是死道友不死贫道。