Wim Vandebroeck

AI Security Architect | Offensive Security | Sovereign AI. Exploring trusted autonomy and local-first systems for high-stakes environments.

[Wim Vandebroeck @WIM_VDB]

@sudoingX 624gb

[Wim Vandebroeck @WIM_VDB]

@jun_song depends ...

[Wim Vandebroeck @WIM_VDB]

@mervenoyann this is great thanks for pointing this out!

[Wim Vandebroeck @WIM_VDB]

@ivanfioravanti buy what you need to run your stuff locally don't wait prices will only go up... Privacy has its price people sharing their thoughts and ideas with public models a bad idea...

[Wim Vandebroeck @WIM_VDB]

@ivanfioravanti I have a 10gig switch and 10 gig ports one is rtx 6000 pro server other is Mac ultra 512gb. if someone can confirm this would work I will try to set it up

[Wim Vandebroeck @WIM_VDB]

@ivanfioravanti is possible to runs model cross nvidea and Mac ultra?

[Wim Vandebroeck @WIM_VDB]

@ivanfioravanti @NVIDIAAI let me know how it goes ! The 6000 with qwen 122b does 159 tokens/ second insane 4 x faster than Mac ultra .

[Wim Vandebroeck @WIM_VDB]

@ivanfioravanti @pcuenq great numbers don't forget 12 k system vs 50k for cards alone in the nvidea case...

[Wim Vandebroeck @WIM_VDB]

@ivanfioravanti we should create a community brain where we share our findings. So all dev agents can benefit.

[Wim Vandebroeck @WIM_VDB]

@ivanfioravanti just a matter of getting the HW :)

[vx-underground @vxunderground]

Chinese government super computer (allegedly) compromised and (allegedly) 10PB exfiltrated.

The source is CNN.

Something about this story is very strange to me. I've been doing cybersecurity stuff for a long, long time. I'm usually on top of most cybersecurity incidents, whether I discuss it publicly or not, yet I have not heard of this story and I have not seen the moniker "FlamingChina" before.

Furthermore, none of my colleagues have mentioned this compromise to me.

I'm very curious who these cybersecurity experts are who they cite in the article.

I'm also very curious on the 10 PETABYTES of data exfiltrated because they is an unfathomable number.

10PB is 10,000 TB. Even in cold storage that's roughly $43,000/month. If it's "hot storage" you're looking at something like, $150,000/month, that doesn't even include the fees for moving the data which would be ASTRONOMICAL.

Very very strange

[International Cyber Digest @IntCyberDigest]

🚨‼️ An angry researcher just dropped a Windows Defender 0day exploit, he has a message to Microsoft:

"I'm not explaining how this works, yall geniuses can figure it out."

The exploit targets Defender's internal signature update mechanism to achieve local privilege escalation.

[International Cyber Digest @IntCyberDigest]

‼️🇺🇸 Lockheed Martin has allegedly been breached and 375TB of data is being offered for sale on what appears to be a Russian 'Threat Market'.

They've priced the highly confidential data at $598 million...

[Wim Vandebroeck @WIM_VDB]

The sovereign AI cookbook: building a private multi-model platform from scratch with @apple mac ultra.
https://t.co/YdCsSyV5Ze

[Aditya Chordia, CISSP, CIPP/E, CISA @Adi_AIGovSec]

The DarkSword iPhone exploit code just leaked on GitHub. This changes the threat model for everyone, not just Apple users.

Here's what most coverage is missing.

DarkSword was originally a nation-state grade tool — tracked by Google's Threat Intelligence Group since November 2025, used by Russian espionage groups and customers of a Turkish commercial surveillance vendor. This was elite capability reserved for high-value targets.

Now it's on GitHub. Anyone can download it, study it, modify it, and redeploy it.

That's the moment a spyware-grade exploit chain goes from "targeted espionage" to "commodity attack tool." Google themselves warned this is exactly what happens - leaked code gives threat actors a starting point to test, tweak, and iterate.

Three malware families deploy after compromise: GhostBlade, GhostKnife, and GhostSaber. Together they steal data, establish a backdoor for re-entry, and execute code - compressing the entire kill chain into a single click.

But here's the enterprise angle nobody is connecting.
SecurityScorecard's CISO Steve Cobb put it perfectly: once attackers gain credentials on a compromised phone, they're no longer limited to that device. They move into SaaS platforms, cloud environments, and partner systems without needing another exploit.

Now think about how many people use the same iPhone for:
→ Corporate email and Slack
→ AI agent control channels (Telegram, WhatsApp, Discord)
→ Two-factor authentication
→ Cloud storage with synced credentials

A compromised iPhone isn't a phone incident anymore. It's an enterprise access incident. If your CISO is running an OpenClaw agent through Telegram on their Mac and their iPhone connects to the same Telegram account - the phone becomes a lateral entry point to the agent.

This is the second iOS exploit kit disclosure this month. Coruna gave attackers 23 exploits across iOS 13 through 17.2.1. DarkSword covers iOS 18.4 through 18.7. Between them, nearly every iPhone version in the wild has been targeted.

What to do right now:
→ Update to iOS 26.3 immediately - this patches the DarkSword chain
→ Enable Lockdown Mode on any device you can't update
→ If your org allows BYOD, assume unpatched personal devices are compromised
→ Review what enterprise services are accessible from mobile - email, cloud, SSO tokens, AI agent channels
→ Test whether your mobile security controls can actually detect and block these exploit chains, not just in theory

The pattern is clear: nation-state exploit tools are leaking faster than organisations can patch. DarkSword is public now. The window between "elite capability" and "commodity attack" just collapsed.

Patch today. Not tomorrow.

More Info: https://t.co/U4G4e1IIkL

[Clément Pillette @ClementPillette]

I've been comparing two setups running Qwen3.5-397B-A17B at full 262K context:

🖥 Mac Studio M3 Ultra (512GB) — €14,500
⚙️ Custom workstation, 4× RTX PRO 6000 (384GB VRAM) — €45,000

Results:
• Workstation: 46.9 tok/s, 1,100W, 51 dBA
• Mac Studio: 35 tok/s, 120W, ~15 dBA

The Mac is 6.7× more energy-efficient per token. Over 3 years, the TCO gap is nearly €40K.

I have never been a Mac guy, but I have to admit that the Mac Studio is currently the most attractive hardware for running local AI agents.

[David Hendrickson @TeksEdge]

🚀 I just hosted MiniMax-M2.5 UNCOMPRESSED (full BF16, 230B total / 10B active MoE, 457 GB weights) on a single Mac Ultra 512 GB!

Fits with ~50 GB left for OS + modest KV cache (perfect for 4k-16k ctx). Using native MLX — zero swapping, max quality local inference.

16K context window @ 30 tps.

Frontier model at home, no API.

📊 Latest AI Coding Benchmarks

🏆 SWE-Bench Verified
• MiniMax M2.5: 80.2
• MiniMax M2.1: 74
• Claude Opus 4.5: 80.9
• Claude Opus 4.6: 80.8
• Gemini 3 Pro: 78
• GPT-5.2: 80

⚙️ SWE-Bench Pro
• MiniMax M2.5: 55.4
• MiniMax M2.1: 49.7
• Claude Opus 4.5: 56.9
• Claude Opus 4.6: 55.4
• Gemini 3 Pro: 54.1
• GPT-5.2: 55.6

🖥️ Terminal Bench 2
• MiniMax M2.5: 51.7
• MiniMax M2.1: 47.9
• Claude Opus 4.5: 53.4
• Claude Opus 4.6: 55.1
• Gemini 3 Pro: 54
• GPT-5.2: 54

🌍 Multi-SWE-Bench
• MiniMax M2.5: 51.3
• MiniMax M2.1: 47.2
• Claude Opus 4.5: 50
• Claude Opus 4.6: 50.3
• Gemini 3 Pro: 42.7

🌐 SWE-Bench Multilingual
• MiniMax M2.5: 74.1
• MiniMax M2.1: 71.9
• Claude Opus 4.5: 77.5
• Claude Opus 4.6: 77.8
• Gemini 3 Pro: 65
• GPT-5.2: 72

✨ VIBE-Pro (AVG)
• MiniMax M2.5: 54.2
• MiniMax M2.1: 42.4
• Claude Opus 4.5: 55.2
• Claude Opus 4.6: 55.6
• Gemini 3 Pro: 36.9

#MiniMaxM25 #LocalLLM #AppleSilicon

[Hackmanac @H4ckmanac]

🚨Cyber Alert‼️

🇪🇺European Commission

The European Commission disclosed a data breach affecting its mobile device management system, detected on January 30, 2026.

The incident may have exposed staff names and mobile phone numbers, with no evidence of device compromise.

The breach was contained within nine hours and is likely linked to exploitation of Ivanti Endpoint Manager Mobile vulnerabilities.

Source:
https://t.co/lonrNrq2FV

[BleepingComputer @BleepinComputer]

EDR killer tool uses signed kernel driver from forensic software - @billtoulas
https://t.co/bbYxttVbia

https://t.co/bbYxttVbia