
GC @ Flashbots ⚡️🤖, Previously: GC @espressosys @nymproject, Associate Professor @BU_Law @uchastings @utexaslaw, BigLaw @orrick, Engineer @sunmicrosystems @sgi
Joined February 2010
- Tweets10.8K
- Following4.2K
- Followers3.7K
- Likes7.6K
ahmed ghappour ⚡️🤖 retweeted
Your network is decentralised. The people running it still live under laws.
For privacy infrastructure, understanding where state power can reach is part of understanding how resilient a system really is.
At Ethereum Cypherpunk Congress #3, @ghappour of @flashbots explored “The Law at Every Hop: Mapping State Power over Privacy Networks.”
Where do builders need more clarity about their legal exposure?
The Flip, in 12 months on @OpenRouter
Share between open-weight models vs private:
Nov 2025 — 20/80
Nov 2026 — 80/20, and still climbing
ahmed ghappour ⚡️🤖 retweeted
🖥️ The Mac you're not using could be serving AI models to other people, and @0xkydo showed us how @darkbloomai already does it: 1,311 Macs in 59 countries handled 151M requests in 30 days, using Apple's Secure Enclave and App Attest to hide your prompt from the Mac owner ⬇️
ahmed ghappour ⚡️🤖 retweeted
No.
1. In the presence of superhuman bug finding machines, humans cannot write safe code without extreme levels of care and planning.
2. That care has almost never been applied to the critical systems in our lives. Not because software engineers don't care, but we aren't really engineers, not in the way civil engineers are engineers. The title is stolen valor.
3. Even careful, experienced "engineers" have difficulty writing code that is safe against modern models in C/C++ and other low-level languages.
4. We are finding that CVEs are overrated; huge numbers of organizations are being owned by custom vulns in the 100s of billions of lines of non-standard enterprise code out there or misconfigurations in heterogeneous systems, which LLMs can find and write exploits for live.
5. @dinodaizovi is right, we need architectures for which one RCE is not game-over and that buy us time for AI-driven response.
6. Everything is going to have to be re-written by AI, in type/memory safe languages, with minimal libraries, using formal methods, contracts, solvers, etc...
The vulnpocalypse is caused by security vendors.
Their inability to convey risk accurately, due to either incompetence or ignorance, causes tremendous damage and stress on responding teams.
There is a lot of weight associated with vulnerability announcements. We can no longer trust many security vendors’ blog posts, since their purpose is marketing & fanfare, not accuracy.
Exploitability needs to be demonstrated, in realistic environments, and its likelihood accurately conveyed. So that responders can make informed decisions on what to prioritize. But instead we are seeing fake PoC videos set in contrived & controlled lab environments, that give an illusion of exploitability when it’s unrealistic in any real-world scenario.
And vendors now try to show off who found it first, rush to send you marketing emails and call you without your consent; to sell you snake oil they pulled out of their ass two hours earlier, with no intention on actually making anyone safer, only to meet their quarterly sales quotas.
The core responsibility of a security professional is to accurately communicate risk, and let people make informed decisions. Too many vendors are failing at this, that I cannot consider them security vendors, only snake oil salesmen taking advantage of people in vulnerable positions due to their lies.
All better now chill people
More news, this time actually cyber-related:
The guys at @calif_io who discovered last month's PHP bug used AI to find a potentially even more severe bug in Java, the programming language used by estimated 90 percent of Fortune 500 companies. Patch just released by Oracle.
1
178
ahmed ghappour ⚡️🤖 retweeted
Notetakers hear every word of your call.
Conclave for Mac takes meeting notes for Zoom and Google Meet, from an app on your Mac. No bot joins the call, and names come from the meeting itself. Video attached.
Built at NEAR HACKATHON @NEARProtocol #NEARHackathon
What NEAR does here:
Transcription runs on NEAR AI's private inference inside a TEE. NEAR AI publishes an attestation for each model it runs for us. Conclave checks the transcription model's and the notes model's, for every meeting and live on our trust page, and shows the result as received.
Every finished meeting gets a receipt. The receipt's seal key id should match the one on app.conclavehq.org/trust. NEAR AI says it signs each transcription inside its enclave but does not yet expose that signature, so no per-request proof is shown.
How the audio moves:
During a recording, Conclave for Mac sends the audio in 15-second pieces to Conclave's server, which runs in a confidential VM on Phala. Each piece is stored encrypted while the meeting is processed.
Try it on macOS 14.2+: download at app.conclavehq.org/download, sign in with Google, record a call. 3 hours free every month. Headphones give the cleanest transcript.
Code: gitlab.com/conclavehq/mac
What would you need to see before you let a notetaker into a client call?
😮 evidence suggests Chinese AI labs using Nvidia chips banned from export to China
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
2
6
314
ahmed ghappour ⚡️🤖 retweeted
congratulations on the launch, excited to be working with @LYSIOS on the scaling trust arena!
🚨 LIFE UPDATE: I’m joining a lab! 🤗
oh, did I say joining? I meant starting.
and did I say lab? well… that would be a biiit of an understatement 😉
INTRODUCING: @LYSIOS ⛓️💥
nine heads. zero fucks. unreasonable ambition.
a Superorg for the Age of Superintelligence.
ahmed ghappour ⚡️🤖 retweeted
whats more cypherpunk than a human exercising their right to privacy?
their agent swarm exercising that same right, across thousands and thousands of agents, until there's a network of shade big enough for anyone to blend into ⬇️

ALT A dark low-poly forest at night. Six searchlights converge on a small gold cube labelled your agent, standing in a clearing inside a crosshair. Each searchlight is labelled with a watcher: frontier lab, inference router, bot wall / CDN, search engine, internet provider, app telemetry. Title: ShadeNet, cover for local agents, shadenet.xyz.
ahmed ghappour ⚡️🤖 retweeted
TRACKED CHANGES: Law firm innovation chiefs left a private meeting with Anthropic unimpressed, per Bloomberg Law. “They need to wake up,” one said, citing two major dealbreakers: no hands-on training support and a minimum spend requirement just to negotiate custom privacy terms.
ahmed ghappour ⚡️🤖 retweeted
Only an AI company can announce what is ultimately a basic set of controls that should have been implemented years ago, and be praised by people who have never worked on complex systems as if this was groundbreaking technology.
We built Control Room, a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise.
It connects every research inference workload, including agentic sampling for model training and evaluation, data processing, and more. This works together with our misalignment monitoring systems, and halts model output within 35 seconds of a decision to intervene to stop or pause a job.
Control Room is just one part of our holistic approach for safe frontier model training and deployment.
openai.com/index/control-roo…
Legends 🫶🏽✊🏽
Enjoyed hanging with @_wjth @nicksvyaznoy @RogerDingledine @ghappour and @misaakidis in Mumbai thanks to @EFDevcon and @web3privacy.
You may think projects like Swarm, Tor and Nym compete, but we are united to defend user privacy!!
5
217
ahmed ghappour ⚡️🤖 retweeted
The French gov has officially attributed the supermarket fridge hacks to Iranian actors
diplomatie.gouv.fr/fr/presse…
ahmed ghappour ⚡️🤖 retweeted
Cypherpunk isn't just code.
Sometimes it's a protest song.
Yalgaar Sanskrutik Manch opened Ethereum Cypherpunk Congress 3 in Mumbai with the spirit of resistance.
An anti-caste art collective fighting oppression through music, reminding us why freedom from surveillance matters for dissent.
This is what The Ancestral Fork sounds like.
Proof of exceptionally confident hand-waving at my first ever applied cryptography talk. Ty @EFDevcon 8!
👀 ✊🏽
Replying to @OrinKerr
As regular readers know, I've been banging the drum for why TOS do *not* eliminate Fourth Amendment rights for years. My 2024 article on this is here: papers.ssrn.com/sol3/papers.…
ahmed ghappour ⚡️🤖 retweeted
Even a private transaction can reveal who sent it as access layer traffic can leak network metadata. Nym's SDK runs the mixnet in the browser with no install and no config, so wallets and dapps can tunnel their traffic through it and hide who's sending.
At DEVCON India @_wjth "The Nym Mixnet works today. With over 700 Independent node operators. "
Wallets can intergrate the Nym Mixnet into their wallet giving users strong network security seamlessly via our SDK.
👀
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
1
2
159






























