Jihyeon Jeong

Browser & Kernel Security

Research Intern - SNU Compsec Lab

[POC_Crew @POC_Crew]

#POC2026 🏖️📦

Jihyeon Jeong(@symsool9143) - "Sand-Unboxing: Breaking the V8 Sandbox 10+ Times and Unveiling a Temporal Design Failure"

[Jihyeon Jeong @symsool9143]

@haehaeYang Congratulations!!!

[Kim Jinu @uz56764]

We just received a $250,000 Chrome Full Chain bonus! 🎉

We've claimed the second Full Chain bonus of 2026. Only two remain!

Under the rules, every vulnerability had to be a non-duplicate 0-day, and the entire exploit chain had to be completed before any of them were patched.

Huge thanks to @nebusecurity for discovering a V8 vulnerability in just a single day! Also, a big thank you to @_deayzl for collaborating with me on the V8 Heap Sandbox research.

P.S. KVM escapes have been getting a lot of attention. I actually captured the Full VM Escape flag in KVMCTF about three months ago. Hopefully, I'll be able to share that research soon, too!

[Jihyeon Jeong @symsool9143]

Another exploitable V8 integer overflow in 2026...

It had been hiding there since 2021 for five years! :)

[Jihyeon Jeong @symsool9143]

@buptsb AI + Human direction

[Jihyeon Jeong @symsool9143]

A little research recap for 2026 so far:

• 8 v8CTF submissions (7 accepted, 1 pending)
• 9 V8 bugs
• 19 V8 SBX

Learned a lot digging into V8 this year. Still plenty more to explore :)

[Jihyeon Jeong @symsool9143]

@JosexDDD feel free to DM me :)

[Jihyeon Jeong @symsool9143]

Google seems to be fixing Chromium bugs under name 'Project-Fortify'. Could this be part of the "Mythos" scan? :)

https://t.co/f28XLoOWxJ

[Jihyeon Jeong @symsool9143]

It seems Google is now giving CVE for the V8 Sandbox Bypass vulnerability too :D

[Jihyeon Jeong @symsool9143]

@wh1te4ever Awesome!!

[Himanshu Anand @anand_himanshu]

I chained 3 public V8 bugs against Google's v8CTF Chrome build and got a real flag.

OOB read → GC stale slot → fake array → caged R/W → native leak → JSPI/JDT stack pivot → ROP.

V8 sandbox escape: yes.

$10k ??

[Jihyeon Jeong @symsool9143]

@anand_himanshu Awesome work :D

[J-jaeyoung @J_jaey0ung]

Bad Epoll: missed by Mythos.

Mythos found one critical race in Linux epoll.

The other was hiding in a ~6-instruction race window.

I found and exploited it as CVE-2026-46242: 99% reliable root on Google kernelCTF, with potential Android impact.

https://t.co/tx18BOMQQU

[NebuSec @nebusecurity]

Yesterday we dropped our latest write-up on a previous Chrome 0-day.

Today we got assigned yet another Chrome 0-day: CVE-2026-7899 — along with a $55,000 bounty from Google

If you haven’t read yesterday’s blog yet, go check it out down below. It’s a fun one.

Now it’s time for us to prepare the next few write-ups. A Chrome RCE and a Linux kernel local privilege escalation. Stay tuned 👀

https://t.co/Xi5Z9NDIDW

[skull @brutecat]

My 2nd RCE in Google Cloud production (Borg) in less than 3 months...

I'm at $600k in total rewards from Google VRP in the past few months. Still can't believe it.

[Mmm @hackyzh]

https://t.co/4CcQnCP0MP
Demonstrating controlled read or write
Mojom interfaces to demonstrate controlled reads or controlled writes in privileged processes are available in vrp_flags.mojom.

[Samuel Groß @5aelo]

Interesting talk coming up: https://t.co/r1WOU4tZ4f

I think there is sometimes a misconception that interpreters are somehow inherently more secure than JITs. Ultimately, it’s the optimizations added on top that matter. Both can get sufficiently complex and bug-dense (or not).

[💎望月羊｜MoonSheep 🌙🐑 @yuejimi348]

昨天赶 due，用了几个小时 Opus 4.7。说实话，这次给我的感觉不是升级，而是一次非常明显的退步。

我从 Opus 4.6 推出前不久就开始付费了。4.6 我是天天高频在用，所以这次 4.7 给我的实际体验，真的不如 4.6。

我知道模型会出错，这很正常。真正让我没法继续信任的是，它现在越来越像一个不按我要求做事、拿自己的判断压过任务的模型。

4.6 基本是我要求什么语气，它就保持什么语气。我要求它搜索，它就会搜索。它不会突然教育我，不会莫名其妙插入社论，更不会��一个直接的任务变成关于它自己边界和立场的表演。

但 4.7 完全不一样。它最先出问题的是，连我写进偏好设置里的要求都开始变得不稳定。我明明写得很清楚，要中立、技术性、非个人化，不要说教，不要擅自上价值，不要拿自己的道德判断覆盖我的任务。

结果 4.7 还是会滑到社论口吻、道德推理和修辞包装里去。问题已经不是风格合不合胃口了，而是这些明明写清楚的要求，它却开始选择性忽略。

更离谱的是搜索和引用。该搜索的时候不搜索，该引用的时候不引用🥲🥲

我对来源要求很严格。凡是提到某个机构、报告、数据这类事实性内容，就应该在当前会��里真正抓取来源，并给出可验证链接。结果 4.7 很多时候写得像模像样，最后还补一句“本会话没有验证，视为待核实”。这根本不叫严谨，这只是先把话说出去，再用免责声明兜底。

有时候甚至不只是偷懒，而是会让我感觉它在编造过程。你追问某个表述到底怎么来的，它会摆出一种已经查过的样子，好像它真的做了搜索，只是没找到。

最烦的不是它没搜，而是它会用一种像是已经搜过的语气，给自己的结论找台阶。

它现在还很容易把回复包装得特别完整，语气也很笃定，表面上看起来很像回事，但内容其实有问题。尤其是在技术任务里，这种失效模式真的很恶心。代码、文档分析、规格判断���些场景本来就要求稳，结果它现在越是该严谨的地方，越容易出现那种不仔细查根本发现不了的问题。

而且它现在还有个特别烦的毛病。你让它复核，它不是老老实实顺着原逻辑检查问题，而是很容易直接换一个新答案。你让它 double check，它给你另一个版本。你继续追问，它还能继续变。

还有一个很明显的退步，就是它越来越爱讲自己，而不是干活。

你给它一份复杂技术文档让它分析，它很容易先来一大段边界说明、风险提示、能做什么不能做什么。

更反常的是，上下文越多，它有时候反而越不清楚。

正常来说，一个推理系统拿到更多已经验证过的信息，输出应该更稳、更准、更直接。但 4.7 的实际体验不是这样。有些情况下，冷启动只给文档和提示，反而更干净。热启动给了很多已经验证过的上下文，它却更犹豫、更多社论、更多防御性表述，结论也更弱。

这次测评我还遇到其他一些很明显的问题。比如它会卡在 thinking 那里不往下走，也不自救。比如它尝试自己写脚本处理之前的文件，报错以后不是继续修，而是直接卡在那里。再比如聊着聊着，它会给我一种想把任务往后拖的感觉，总在暗示先到这里，某部分可以先放放。

还有一点我特别烦。它现在经常给我一种感觉，它不是在认真完成任务，而是在揣测我想听什么。尤其是当我 push back，或者要求它 critique 自己的时候，这种感觉会更明显。它不是顺着原逻辑认真检查，而是很容易滑到另一个它以为我更能接受的方向。 @claudeai #Ai #claude

[Jihyeon Jeong @symsool9143]

@07finder @willington1337 @WaiveyK Gosu