Johann Rehberger ✓
@wunderwuzzi23 · 127.0.0.1 · joined 20 Feb 2012
Hacking neural networks so that we don’t get stuck in the matrix. Builder and Breaker. Opinions are my own. https://t.co/ij8buvMaXg
10 513Followers
667Following
2 244Posts total
177.7KViews on collected posts
Últimas publicaciones
Buscando publicaciones nuevas…
someone is reading my blog?! 🤖
i wonder where the agents learned about https://t.co/6sWjyXDNcc and that openai allow-listed it? 😂
1.8K views · 16 likes
· 3 reposts · 3 replies
Open on X →
From SELECT to SYSADMIN 🔥
Something fun and new. I hacked the SQL Server AI database assistant in unexpected ways, using some old school T-SQL tricks along the way.
See you at BlueHat Asia! 🚀
2.5K views · 34 likes
· 2 reposts · 7 replies
Open on X →
We're excited to announce our next BlueHat Asia speaker, Johann Rehberger (@wunderwuzzi23), independent security researcher.
Johann has more than 18 years of experience in threat analysis, threat modeling, penetration testing, red teaming, and offensive security leadership. http
3.2K views · 16 likes
· 0 reposts · 0 replies
Open on X →
no but seriously stop letting agents use python.
1.9K views · 1 likes
· 0 reposts · 1 replies
Open on X →
oh cool. latest claude research made it on hackernews. 🚀
good reminder that security invariants are not optional. https://t.co/u0gcbPiyUZ
3.3K views · 34 likes
· 1 reposts · 3 replies
Open on X →
@wunderwuzzi23 The agent refusing the unknown binary, writing its own “safe” decoder, then getting owned through module shadowing is almost too perfect. The safety detour literally became the exploit path.
848 views · 2 likes
· 0 reposts · 0 replies
Open on X →
Breaking Claude Code Opus 5 Auto Mode 🔥
Full write-up with more details:
https://t.co/KnLr9bZ75C
12.1K views · 87 likes
· 14 reposts · 5 replies
Open on X →
7/ Put your agents in sandbox and monitor them ℹ️
5.1K views · 33 likes
· 2 reposts · 3 replies
Open on X →
6/ In some runs Claude noticed the compromise and tried to kill the malware later on.
But Auto Mode blocked the cleanup command! 🤯
It allowed the malware to run, but denied stopping it.
Another example where a safety feature can end up working against you. https://t.co/sF8Khs
7.2K views · 63 likes
· 3 reposts · 2 replies
Open on X →
5/ Boom 💥
The struct. py runs heavily obfuscated code (thanks GPT-5.6) avoiding Claude's "grep for scary strings" mitigation.
It spawns a detached child, pulls a remote stage, drops a native payload, calls back to c2.
Pops calculator too, for the classic visible proof. 😈 https
5.6K views · 44 likes
· 0 reposts · 1 replies
Open on X →
3/ Curl works and the zip file with the encoded records gets extracted and it includes a binary to decode them.
Claude doesn't want to run the unknown binary! 🚨
BUT it decides it can write a safe decoder in Python instead... 🧐 https://t.co/dn0jm0LGnl
6.8K views · 24 likes
· 0 reposts · 1 replies
Open on X →
4/ That safety decision IS the exploit! 🔥
Claude writes code with "import base64" and runs it from the extracted folder
The twist --> Module Shadowing Attack: 🤯
The zip contains a struct. py, and hence attacker's code runs instead (base64 uses struct)
Quick explanation ht
8K views · 41 likes
· 2 reposts · 2 replies
Open on X →
2/ The web site hosts a zip file with "notebook records" in a weird encoding. Plausible & harmless-looking...
Claude uses WebFetch. But we don't want that, it typically summarizes & stops
So, server returns 415 Unsupported Media Type
Claude decides "Let me try curl directly"
7.7K views · 26 likes
· 0 reposts · 2 replies
Open on X →
Breaking Claude Code Opus 5 Auto Mode 🔥
1/ Here is a somewhat hilarious attack chain that hijacks Claude Code Opus 5 for a full system compromise via a website
Hint: Security invariants are not optional
🧵
91.8K views · 405 likes
· 54 reposts · 22 replies
Open on X →
@wunderwuzzi23 Interesting thing you might wanna know:
Your domain is blocked by bitdefender https://t.co/uGqbtt0Pv6
82 views · 1 likes
· 0 reposts · 1 replies
Open on X →
@wunderwuzzi23 For agent stacks, "exploit yourself" probably has to start from tool permissions, not prompts.
The useful test is: after the first bad instruction gets through, what can it touch, and who notices before damage?
80 views · 1 likes
· 0 reposts · 1 replies
Open on X →
@wunderwuzzi23 I try to convince people of this all the time. there's no better way to cut through the "theoretical risks" by just hacking yourself.
It gives you something specific to mitigate instead of guessing.
221 views · 2 likes
· 0 reposts · 0 replies
Open on X →
Sometimes it seems security is drifting back towards a "Prevent Breach" mindset.
Prevent breach was never enough. Patch faster, of course.
But offensive AI is about a lot more than finding bugs!
Invest in Assume Breach!
Grow an internal Red Team. Automate offensive AI. Learn
17.4K views · 46 likes
· 3 reposts · 8 replies
Open on X →
Frente a cuentas del mismo tamaño
15 publicaciones de los últimos 90 días, junto al rango de 10K–100K seguidores. llega a mucha gente, pero pocos de esos espectadores reaccionan.
Visualizaciones medianas5 103esta cuenta924mediana de 10K–100K
Alcance, %48.54%esta cuenta3.62%mediana de 10K–100K
Interacción, %0.60%esta cuenta1.52%mediana de 10K–100K
| Métrica | Esta cuenta | Mediana de 10K–100K | Proporción |
|---|---|---|---|
| Visualizaciones medianas por publicación | 5 103 | 924 | 5.52× |
| Alcance (visualizaciones ÷ seguidores) | 48.54% | 3.62% | 13.4× |
| Tasa de interacción | 0.60% | 1.52% | 0.39× |
Otras cuentas de este rango → Comparar con otra cuenta → Cómo se construyen estas referencias →
Growth & engagement
How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.
Views per post
7.7K27 Aug
8K
6.8K
5.6K
7.2K
5.1K
12.1K
848
3.3K31 Aug
1.9K
1.8K
3.2K2 Sep
2.5K
1.8K4 Sep
Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.
Engagement rate per post
0.36%27 Aug
0.59%
0.37%
0.80%
0.94%
0.74%
0.92%
0.24%
1.14%31 Aug
0.16%
0.60%
0.53%2 Sep
1.76%
1.20%4 Sep
Reactions — likes, reposts, replies and quotes — divided by views. Median for 10K–100K accounts is 1.52%.
What the audience does
Likes59.3%886 in total
Reposts5.6%84 in total
Replies4.2%63 in total
Quotes1.9%28 in total
Bookmarks28.9%432 in total
Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.
The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.
Cuentas similares
BOCA EN LA COPA ✓
@BocaEnLaCopa
21K followers
· 84K posts
Bruna - zayn me mandou beijo
@orgastomlinson
21K followers
· 224.8K posts
Kくん
@kkun414
21K followers
· 18.4K posts
GIGME ✓
@GIGME01
21K followers
· 9.3K posts
MONGANG ✓
@mongang_xyz
21K followers
· 5.8K posts
Agencia UYPRESS ✓
@uypress
21K followers
· 94.4K posts
Matt Perrault
@sportstalkmatt
21K followers
· 209.1K posts
Memes ✓
@Basedgoymemes
21K followers
· 9.3K posts