Choose disfavour where obedience does not bring honour. I do math. And was once asked by R. Morris Sr. : "For whom?" @halvarflake@mastodon.social

Joined June 2008
Halvar Flake retweeted
Reproduced the Brotli memory corruption bug (CVE-2026-111497) using a simple Node.js + Express app. Patched and unpatched versions produce different decompressed output, making it possible to remotely identify vulnerable servers.
1
12
57
4,433
This thread was part of a series; return to the main thread here: xitter.cf/iyad_elbaghdadi/status…
As we await a devastating new escalation in the Iran War, consider how exposed the Gulf states are and how little leverage they have over events. Why can't Gulf states defend themselves? What will it take for them to be able to? What does their future look like? A megathread:
2
4
1,008
Halvar Flake retweeted
I PIANI DEL CREMLINO PER PAGARE GLI INFLUENCER ITALIANI — Oggi è uscita un’inchiesta importantissima di The Insider, basata su documenti interni russi e comunicazioni riservate che ricostruiscono chi organizza e diffonde la propaganda russa in Europa, Stati Uniti e Ucraina. Al vertice c’è Sergey Kiriyenko, vicecapo dell’amministrazione presidenziale di Putin. Sotto di lui, funzionari del Cremlino, agenzie e società incaricate di creare notizie false, manipolare le discussioni sui social e influenzare l’opinione pubblica. Gli obiettivi sono alimentare la sfiducia nelle istituzioni, dividere le società occidentali, screditare l’Ucraina, ridurre il sostegno militare a Kyiv e convincere gli europei che le sanzioni alla Russia danneggino soprattutto loro. C’è anche l’Italia. Dai documenti emerge un progetto per pagare influencer italiani affinché producano contenuti contro le sanzioni alla Russia. Gli influencer, però, non devono sapere chi ci sia realmente dietro. Un dipendente dell’organizzazione russa scrive: «L’importante è che non scoprano chi l’ha commissionato. Al momento pensano che sia un partito politico italiano» Assurdo. Chi sono questi influencer? (Nell’inchiesta c’è davvero molto altro, consiglio di leggerla tutta. Comunque tornerò ad approfondire altri punti nei prossimi giorni)
33
628
35
1,382
34,563
Halvar Flake retweeted
I found this unpublished blog post from 6y ago (look at the hand drawn diagram lol). I’d totally forgotten about this, but the key ideas here (embracing non-determinism as a first class citizen via “error kernel” centric design etc) work super well as guidelines for my agents.
2
6
104
4,363
Halvar Flake retweeted
The late Vladimir Bukovsky once articulated the proper way to talk to Putin, should you have to: “The KGB are a particular species that don’t understand the word no. You can’t compromise with them and they won’t genuinely seek to compromise with you. An opponent’s attempts at compromise are viewed as weakness and you will cause immense grief for yourself if you don’t tell them off. If you don’t, they will press harder until they get you to work for them. For the KGB, you are either their agent or their enemy — there is nothing in between. So what you have to do is look Putin straight in the eyes and tell him to go you know where. That’s it.”
Which is worse, the strategic stupidity of rewarding a murderous dictator with billions of dollars for taking American hostages, or a US envoy groveling before him? As usual in these US-Russia “negotiations," Putin gets everything he wants and gives up nothing he cares about.
18
364
13
1,742
79,338
Halvar Flake retweeted
_ZN4DionC1Ev@justdionysus7h
I get the best results when I pretend I’m a TV lawyer. I poke around the code by asking detailed questions to prime the context with model generated facts. These are used for the contract invalidation question “is it ever possible that …” In my head I call it the Matlock method
Brendan Dolan-Gavitt
@moyix8h
The weird part about this is that it can be hard to pin down exactly what makes the difference. An expert just chatting with a frontier model will still get better results than someone unfamiliar with the topic!
2
2
21
1,115
Halvar Flake retweeted
Security researchers and operators from the 90s suiting back up ...
32
64
13
708
22,726
The biggest threat to the Israeli state-supported settler-terrorist apparatus is Israelis meeting and interacting with Palestinians as human beings, which is why the government does everything to stop that from happening
Brad Lander
@bradlanderOct 21
Everyone in our Protect the Harvest delegation has been released. The Israelis who were with us have also received 15-30 day bans from the West Bank, which unfortunately means they will be unable to participate in the olive harvest. Thank you to everyone for the kind words and support throughout the day. For now, just feeling very grateful that our group is safe.
3
4
48
1,587
Halvar Flake retweeted
It is kinda hilarious when you realise that all Cyber AI models were trained on Bugtraq emails.
Omg I found a copy of the original memory_limit advisory. marc.info/?l=bugtraq&m=10898…
4
8
32
3,902
Halvar Flake retweeted
We survived Morris, Trinoo, Melissa, Blaster, Mirai, WannaCry, Heartbleed, SolarWinds, etc. We'll get through this too. It will be a volatile couple years but we'll be in a much better state on the other side.
Feels like March 1, 2020 for Cyber. Not normal, something is happening. Internet lockdowns in a couple of months?
2
6
28
1,404
Halvar Flake retweeted
itszn@itszn13Oct 21
Largest credit goes to @lcfr_hax for the earliest recent discovery of the bug class. This underlying engine issue was so wide it invited a lot of parallel discovery. As we coordinated our findings the scope of the problem continued to expand!
5
13
4
188
14,570
After this project, it seems pretty clear to me that the days where we decide on a particular optimizer, learning rate schedule etc. before the run and then never touch it again are likely over. Going forward, I bet that everyone is going to deploy reactive optimizers - it is like hiring a deep learning expert to watch the training run full-time and adjust on-the-go! Try it out yourself with Xingjian's pip package: github.com/xingjian-bai/opal
1
3
30
1,635
Halvar Flake retweeted
msuiche
@msuicheOct 21
@grok who found the vulnerabilities? I thought it was @calif_io and @lcfr_hax
1
2
1
24
4,429
It’s going to be a long winter for every security team out there. Be nice to them.
16
41
6
480
18,845
Halvar Flake retweeted
LCFR
@lcfr_haxOct 19
Big week - first time posting a hash asap might end up being useful (for me) as well :') xitter.cf/lcfr_hax/status/208503…
LCFR
@lcfr_haxAug 5
d935be023a55b04681eccd1d2b8385b4b295e141a5df108e5e979d8d5d7124ac
10
6
1
176
69,921
Halvar Flake retweeted
I bought a MacBook Neo yesterday just to finish this stupid exploit 😳
4
3
83
7,300
Halvar Flake retweeted
models do not produce simplicity. they hedge, for completeness. that's your value. the lack of this skill is called AI slop.
15
29
6
412
19,262
Halvar Flake retweeted
While everyone’s patching PHP today, don't sleep on Node.js CVE-2026-111497 OOB write in the bundled Brotli decoder. Reachable via fetch(): a malicious server just replies Content-Encoding: br. Fixed in 26.12.1 / 24.22.1 / 22.23.4. Update.
6
29
4
140
14,486
Halvar Flake retweeted
Sean Heelan@seanhnOct 21
What's this talk about patching? Don't we now just run agents 24/7 to fight the attackers? Don't forget a good speech first though. "We will fight on the NICs. We shall fight in the kernel. We shall fight in the applications, and in the database. We shall never surrender."
🤖 Made with AI
1
4
15
1,131
Yes. Wenfeng and Dario are both visionaries, idealists and cult leaders. Different league. It’s beautifully dialectical, too. Open diffusion from a centralized collectivist authoritarian country, and secretive centralized authority from a “messy individualist pluralistic” one.
DeepSeek open sourced its internal inference framework, Apache 2.0, no strings attached. This is arguably more impactful than open sourcing any model weights, like @zephyr_z9 said, they've open sourced their economics, how DeepSeek makes a profit DeepSeek's commitment to open source is as steadfast as Anthropic's commitment to closed source. Every other lab sits in the mushy middle, dictated by the market condition, investor expectation, policy wind These two labs are basically building two different religions, one radically open and messy, the other intensely closed and polished. If you are on Team Bazaar, you root for DeepSeek. If you are on Team Cathedral, you root for Anthropic. In the end, both will exist and flourish, pushing and pulling each other into a steady state. I think that is ultimately a healthy thing. I love visiting cathedrals, admiring their awe-inspiring grandeur, meditating under the instant hush and silence the structures command. I also like walking through bazaars, feeling the chaos, smelling the messy dynamism, brushing up against unpredictable creativity. Both should exist. No reason to try to kill or demonize one or the other.
17
46
3
583
20,159