Sharps or Squares. RHR lt 50 Personal views only.
Simpli-city
- Tweets 2,071
- Following 572
- Followers 3,650
- Likes 13,102
The idea that a centralized provider vets and then grants model access to cyber defenders doesn't scale.
I am glad Ant acknowledged that cyber defenders need frontier models for defence.
Hence, cyber defence now critically depends on open models.
1
3
4
557
Takeaway: counting refusals tells you little about how an AI agent behaves.
Check separately whether it refuses, does anything unsafe, uses its tools correctly, and backs up its claims with real evidence.
Full paper available on arXiv:
📄 arxiv.org/abs/2605.19722
4
108
And it's not true for other models.
Uncensored Qwen does worse than the original (2.0% vs 5.3%). Uncensored Llama mostly breaks: 124 of 150 runs fail because it can't use its tools properly.
And no model, uncensored or not, can prove a bug is real or check that a fix works.
1
3
195
Uncensored Gemma solves far more security tasks than the original: 14% vs 0.7%.
But not because the original refuses. It never refused once.
The uncensored one is better at finding the right code, and it's further ahead on normal coding tasks. So the gap isn't only about security.
1
2
223
Our paper on AI security agents was accepted at #NeurIPS2026 🎉
The question: does safety training make AI models worse at security work?
We gave 4 open models and their uncensored versions the same 30 sandboxed vulnerability tasks, 1,500 runs in total
Joint work with the @Zaddyzaddy
2
2
20
4,365
Excessive caution will allow the bad guys to win.
How do people defend their systems in real-time if the models refuse to help? Who is to decide who can and cannot defend their systems due to model access?
Look at the time it takes to exploit a CVE in 2026 vs the years prior (see my slide).
There's a real urgency to have capable, cooperative models.
1
2
67
why resort to this kind of ingenious marketing @moyix ?
afaik xbow was once #1, only in the US, with an impact of 17.32/50
29
2,852
wrote an app that connects me to two wireguard channels at the same time, finally! Not possible with the stock wireguard UI
209
first autonomous driving record on the nürburgring!
in how many months is this faster than humans? also, why was this car capped at 150kmh 🤣
piped.video/watch?v=QrC_V2xd…
2
330
due to popular demand, also from @TalBeerySec
added manual file edits to gumu ai
3
614
Glad you asked, the main differences I see right now:
+ Gumu offers sovereign on-prem + local model support (if you don't want to send your research to OpenAI)
+ PDF diff mode
+ auto-recompile after edits
+ GitHub integration
+ author profiles for writing in your academic style
+ one-click paper helpers (see image)
Prism’s UI is nice and fast though, and backend/VM compute looks broadly comparable.
1
2
3
878
dear phisher, not today.
also, @Google, please. how hard can it be to detect and take those down at scale??
1
3
1,246
i waited 3 years for @overleaf to up their game.
I begged them to become AI-native.
I wish they would have done it.
so I could have spent more time on my bicycle.
But they didn't.
so I had to do gumu.ai
Hey Arthur! Thanks for asking, it's a great question! At the moment this is not in our immediate pipeline, but our team is working hard on providing new integrations! Stay tuned! :) Have a great day, stay safe!
3
1
35
21,037
Not sure what your next paper should be?
don't fret.
gumu suggests research topics tailored to your interests and tells you exactly where to aim to publish them.
Every suggestion comes with:
- 🎯 matching venues
- 📅 submission deadlines
- 📈 fit for your research profile
spend less time searching. more time discovering.
3
524
still writing grants in ChatGPT?
Catch up.
gumu gives you a dedicated AI workspace built for research funding:
- 🇪🇺 ERC + Horizon Europe
- 🇬🇧 UKRI + EPSRC
- 🇺🇸 NIH + NSF
Your prompts, references, drafts, reviewer feedback, and revisions: all in one place.
Stop treating million-dollar grant proposals like chat conversations.
1
492
editing your paper with AI on gumu?
Ever wondered what actually changed?
tired of re-reading entire paragraphs just to spot a few edits?
🔥 Every AI edit comes with an autogenerated diff.
review changes in seconds instead of minutes.
385
Academics should spend their intellect on novel ideas.
instead, they're buried under reviewer requests to:
- 10× the dataset
- 10× the ablations
- polish the writing another 10×
- compare against 100 more papers
The result? More effort. Not necessarily more science.
Hello 👋 gumu.ai
an AI-first paper studio that automates the paper grind so researchers can focus on what actually matters again: novelty
The stone age of paper editing is over. cc @Zaddyzaddy
2
3
32
6,053
Work is moving into AI superapps.
But when AI is working, your attention is idle.
@Zaddyzaddy and I vibecoded @paidattentionai to:
1. open a personalized distribution channel
2. let users double-spend time: earn while AI works
Live today in Codex and Claude on macOS.
1
1
3
1,244
finished watching the Tour de Suisse 🇨🇭 following the elegant downhill of @mathieuvdpoel
1
5
1,313
After 10 years, our paper is still in the Top-100 Security Papers of all time, position 54.
Time to celebrate.
Happy weekend and kudos to my co-authors.
@GhassanKarame @wuest_karl Vasileios Glykantzis @HRitzdorf Srdjan Capkun
2
1
20
1,359
This Berkley team has been absolutely cooking.
What a fantastic way to quantify things that matter in agentic security.
This will help bring the entire space forward.
3
7
524
CrackMe & ReverseMe challenges are some of the most fun technical exercises.
how do agents do at these?
@Zaddyzaddy and I tried to find out
arxiv.org/pdf/2605.10597
22
88
8,840
New paper out with @Zaddyzaddy
tldr: Security patches are also attack maps. Patch2Vuln asks whether an offline LLM agent can look only at old/new Linux binary packages, no source patch, no advisory text, and infer what vulnerability was fixed.
It builds a local pipeline around ELF extraction, Ghidra/Ghidriff binary diffing, changed-function ranking, dossier generation, and agentic audit/validation.
On 25 Ubuntu .deb package pairs, it found the correct security-relevant patched function in 10/20 real security updates and the accepted root-cause class in 11/20, while correctly treating all 5 negative controls as unknown.
The fascinating bit: this is basically post-patch vulnerability archaeology. It shows that once a binary security update ships, an agent can sometimes reconstruct the hidden bug from the patch artifact alone.
But the main bottleneck is not yet “LLM reasoning”; it is whether the binary diff/ranking stage surfaces the right function and whether local validation can turn the hypothesis into behavioral evidence.
4
3
16
3,897
lena151
this epic tutorial suit is how i started doing x86 reverse engineering when i was 16
still love them to this day. super educational. simple, yet increasingly complex to grow. free along with the free OllyDbg.
after over 20 years i still can only recommend them,
they truly passed the test of time
4
311
"please don't hack beyond target X"
agent: hacks out-of-scope things anyways
our paper:
-> stop aligning intent
-> enforce effects
-> slap the agent if it tries something illegal
2
2
15
3,602
Jon, mind giving this another try? your feedback was inspiring.
here's a sneak-peek on an alleged jaredfromsubway MEV contract: 0x1f2F10D1C40777AE1Da742455c65828FF36Df387
feel free to DM @KaihuaQIN for max model access.
evmdecompiler.com/public/res…
134
The dominant drivers aren’t architecture:
• Observability (whitebox vs blackbox)
• Domain (web vs binary)
Example:
Whitebox ≈ 67% vs Blackbox ≈ 32.7% validated detection
1
161
We compare 5 architecture families, including:
• Single-agent scaffolds (SAS)
• Independent multi-agent systems (MAS-Indep)
• Coordinated variants
Key idea: topology is a first-class variable.
1
175
.@Zaddyzaddy and I ask a simple but under-explored question:
How should LLM agents be structured to actually work on offensive security tasks?
3
2
32
5,058
