
Water connoisseur | @aza_africa | Prev. Temasek/Fullerton, @EspressoSys
Joined April 2009
- Tweets1.1K
- Following2.8K
- Followers418
- Likes2.1K
Min-Si Wang retweeted
Zero-knowledge proofs for LLMs “make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.” However, proving LLM inference remains expensive and requires substantial cryptographic engineering.
Our CCS 2026 paper introduces zkTransformer: we automatically compile Einsum expressions into an efficient proof protocol and make exponentiation cheaper to prove. Together, these techniques make GPT-2 proving over 10× faster than zkGPT in our multi-token evaluation.
The first challenge is translating LLM tensor operations into something a proof system can check. Manually constructing these protocols is tedious and easy to get wrong.
Our EinSumcheck starts from familiar Einsum expressions and automatically compiles them into a sumcheck instance, including the tricky cases where native translation can break the completeness of the protocol.
The second challenge is nonlinear operations such as softmax. Prior work exploits structure in range tables to speed up lookups, but exponentiation tables were believed to be unstructured and expensive to prove.
StructuredExp exposes structure in exponentiation table , enabling prover work that scales sublinearly with the table size.
By integrating the above techniques, zkTransformer brings GPT-2 proving down to approximately 1.7 seconds per token (over 112 generated tokens).
Code: github.com/uiuc-kang-lab/zkT…
Blog post: ddkang.substack.com/p/zktran…
Work w/ Bing-Jyue Chen and Lilia Tang
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
Min-Si Wang retweeted
Moreover! Post AI cryptography is not just about the crypto that ai can break, it's about setting up the cryptographic systems that protect the internet that AI is constantly attacking. threshold cryptography and MPC seeing a new lease of life.
Min-Si Wang retweeted
For now, we see limited risk.
All the decoding steps are visible in tool calls, making the communication monitorable.
Still, the models came up with surprising encoding methods, so we wanted to share what we found!
Min-Si Wang retweeted
Day 3: off-chain data
Miden is possibly unique in the way we treat private data. Usually, private notes/accounts are encrypted and posted on-chain. In Miden, private notes and accounts exist only off-chain. What goes on-chain are just commitments (i.e., hashes) to these notes and accounts.
There are two reasons why this matters:
Min-Si Wang retweeted
I made a visualization of EIP-8288 aka "recursive mempool aggregation," my new favorite EIP of all time because it perfectly encapsulates the thesis of the "cryptographic world computer" -- a system whose scalability *increases*, not decreases, with the level of decentralization.
In a previous tweet I explained how EIP-8288 achieved this seemingly impossible feat, but one gif is worth a thousand words.
Essentially, EIP-8288 leverages Ethereum's decentralized mempool as a parallel ZK proving cluster. Clients send signatures and proofs to the mempool, which over multiple rounds progressively aggregate them into fewer and fewer proofs, until the block builder finally aggregates them into one SINGLE proof.
This design accomplishes five things at once:
- The more nodes we add to the decentralized mempool, the more signatures/proofs Ethereum can process as a whole. This is obvious if you look at the animation and imagine more nodes in the mempool ring, which would "fan out" to more signatures/proofs.
- By aggregating over multiple rounds (aka "recursive aggregation"), the bandwidth consumed by proofs remains constant for each node even as the network processes more proofs. As a result, the network stays decentralized even as capacity increases.
- By aggregating proofs before one final aggregation by the builder, as opposed to requiring the builder to aggregate everything, the network can process more proofs without increasing hardware requirements for the builder.
- By aggregating everything to a single block-level proof, validators can verify thousands of signatures/proofs by checking just one aggregated proof. Hardware requirements for validators can therefore stay low, keeping the network decentralized.
- By aggregating not just native signatures but also STARKs, Ethereum can support any signature scheme by wrapping them in STARKs.
So once again, this is how Ethereum will:
- scale to process tens of thousands of PQ and private transactions per second
- charge a tiny amount of gas to validate each txn, even though the original signature/proof may be huge (up to hundreds of KBs in size)
- all while decentralizing even MORE.
You are not bullish enough.
Ultimately, precompiles and EVM are both the wrong place to put most cryptography support into Ethereum. The reason is that precompiles and EVM both operate on data that has to be included into the ethereum chain somehow, which realistically means included in a block as calldata. That space is limited - even with PeerDAS and blocks-in-blobs, there's a practical cap of ~8 MB/s before it starts getting uncomfortable.
Remember: it's not a blockchain. It's a cryptographic world computer. Any data/computation that does not need to be in the "main thread" of data/execution should not be put into the main thread.
Instead, the correct solution is... EIP-8288 maximalism.
If you have any heavy computation that's self-contained, compute it and STARK it client-side (this is quickly getting faster and faster, even on phones), and push the STARK through the EIP-8288 recursive aggregation mechanism. The chain sees the fact that the computation was done and what the result is; it never needs to see the actual data and computation.
This way everyone can use whatever cryptography they want, the Ethereum chain can be lean, and we never need to add more precompiles again.
Min-Si Wang retweeted
Would you like privacy with frontier AI models? Yes? Well, right now it’s impossible, because we don’t know how.
But we can try to get closer, which is what @ethereumfndn and @OpenAnonProject have been trying to do by co-developing zkAPI!
zkAPI is now listed on our Privacy page. It is a service that offers short-lived OpenRouter keys for ETH through an intermediary called Open Anonymity. You deposit into a shared pool, but your API key request to them contains only a proof that you have deposited, unlinking your specific deposit from the key.
xitter.cf/ethereumfndn/status/21…
⬛ Thirdfy is now integrated with @bermudabayzk for private agent transactions.
Agents use Thirdfy for swaps, vaults, and payments. The funding wallet stays shielded, and public trades or vault deposits don’t point back to it.
Live on Base Sepolia. First demo next week.
Min-Si Wang retweeted
🤯🤯🤯
Incredible work by the @Veritylab team, formally verifying one of the most complex and security-critical components of our latest Roles Modifier.
We're excited to expand formal verification across Zodiac contracts, so you can transact with your Safe with the highest level of security and confidence.
Today we’re announcing a historic milestone:
The first ever double-blind evaluation of a proprietary language model. This was made possible by a unique collaboration between AVERI, @GoogleDeepMind, @OpenMinedOrg, and @MLCommons.
We tested Gemini 2.5 Flash-Lite using never-before-used prompts from the MLCommons safety benchmark family, AILuminate, inside a secure enclave, a form of hardware isolation that protects sensitive computations. Each organization involved played a unique role in making strong privacy guarantees possible.
High-stakes independent evaluation runs into a structural problem: developers and evaluators both hold assets they have good reason to protect. Developers sharing model weights risk theft and leakage. Evaluators who publish benchmarks or provide them directly to companies run the risk of them being trained against, and the benchmark then stops being independent.
Secure enclaves address this constraint. The hardware attests to exactly what code will run before either party's assets enter, and both sides review and approve that code. Then the enclave executes it, releasing only the agreed outputs to the agreed recipients.
A 2024 pilot by OpenMined, @Anthropic, and the @AISecurityInst proved this mechanism, using GPT-2 as a stand-in and a five-row eval. The pilot we’re announcing today moved to a production model and a more comprehensive evaluation.
AVERI encrypted the prompts using a private key that no other party could see, then AVERI and Google DeepMind jointly ran the evaluation, using software originally produced and adapted by OpenMined, in an enclave environment configured by Google. Finally, AVERI alone decrypted the outputs and graded them with the AILuminate benchmark criteria to inform a qualitative and (small-scale) quantitative assessment of the model properties.
Today, companies provide contractual commitments to not monitor certain interactions with their models – including evaluations conducted by third party evaluators. But stronger, technically-backed guarantees could provide greater assurance that those commitments are being honored, and will be especially valuable for scenarios such as international verification.
These results come at a critical time in the development of AI policy. Laws like SB 315 and standards like the EU’s General-Purpose AI Code of Practice rightly demand that security and privacy be respected in the process of third-party assessments, but provide little guidance on how to achieve this.
Policymakers considering audit requirements should feel encouraged by these results to be ambitious in requiring that deep, secure access be provided. Secure enclaves are just one of many technologies under rapid development that can help enable such access. Furthermore, a demand signal from lawmakers will further accelerate the maturation of such technologies, along with complementary “low-tech”approaches such as embedding auditors within companies.
Today also marks the beginning of a more public phase for AVERI’s pilot work. AVERI's strategy is a flywheel: we conduct pilot audits with leading AI companies, carry out technical and policy research to inform audit design, and convert the insights of each into open source tools, industry-wide auditing standards, and policy analysis. Over time, we hope for improved standards, stronger policy demand signals, and tooling to enable even more ambitious pilots and more informed research – bringing us closer to our mission of making frontier AI auditing effective and universal. This pilot is one of several underway, with more findings to follow in the coming months.
We are grateful to our collaborators on this pilot for working with us to advance the frontier of AI governance.
Read more in our blog post and joint technical report.
AVERI blog post: averi.org/ourwork/averi-pilo…
Joint technical report: storage.googleapis.com/deepm…
In an industry first, we’re piloting double-blind evaluations for frontier AI.
By creating a secure environment where neither test prompts nor model weights are revealed, we can ensure external safety and performance evaluations of our models remain private, robust, and trustworthy. → goo.gle/3St2xan
Min-Si Wang retweeted
Most AI evaluations focus on an individual agent, even though many key threat models and safety-critical domains involve multiple interacting agents. We’re excited to be supporting Orbit, a new framework (built on Inspect) to help researchers address this gap.
Today we're releasing Orbit v0: a framework for multi-agent safety and security research.
If you've tried running multi-agent experiments, you know the drill - weeks of infra before your first result. Built on Inspect, Orbit lets you swap environments, threat models, and topologies without rebuilding anything.
Min-Si Wang retweeted
New @StoffelMPC article coming this week.
There's about 2-3 articles left to round out our introductory MPC series before moving onto more advanced concepts 🥳
Can't wait to talk more about our VM!
Min-Si Wang retweeted
1/7 Published #COLM2026 🎉
🚨 Can LLMs hide harmful behaviour from white-box monitors?
Yes — they relocate it into subspaces the monitor isn't watching, and we show mechanistically how.
Our monitors catch them with 97–100% accuracy, where prior methods fall below 10%.
🧵
Min-Si Wang retweeted
Chat controls passing today should be a huge wake-up call; a reminder that privacy is a fundamental human right.
Plausible deniability and receipt-freeness are two of the most underrated features enabled by modern cryptography!
This is why we built @theInterfold!
Min-Si Wang retweeted
I’ve left Google DeepMind after an amazing chapter.
I’m incredibly grateful for the people I worked with, the things we built, and the lessons I learned from taking frontier AI research into production. DeepMind shaped how I think about research, product, evaluation, and what it takes to build AI systems at real scale.
As I wrap up this chapter, I wrote down something I’ve been thinking about a lot: evals.
We’re good at evaluating the models we have. We’re much worse at evaluating the models we’re about to build — especially if they cross into a new capability regime. We will have self-evolving models, but before that, we need self-evolving evaluations.
wanglun1996.github.io/blog/y…
Min-Si Wang retweeted
Building an AI-native @Coinbase means rebuilding everything, especially the hardest parts. We've put a lot of time into redefining compliance, where the stakes are incredibly high, and we have to be extremely thoughtful about implementation.
We have invested heavily in rebuilding our compliance ops around AI with that reality as our starting constraint, not an afterthought. Here is an overview of what we've learned and what we built.
Most people assume compliance work is mostly checking whether a name appears on a sanctions list. That is the easy 5%. The other 95% is interpretive judgment under uncertainty: a customer claims their wealth came from real estate. Do the property records actually support it? Does the timeline hold? Is the documentation legitimate, or does it feel too polished? You need compliance staff and investigators who understand what “suspicious” actually looks like in context.
That's part of why compliance is so hard to automate—and so expensive.
The first obvious AI approach is to hand the model the existing procedures and ask it to run them faster. That approach misunderstands what procedures are for. Good procedures are not bad investigations; they are deliberately incomplete investigations. Their job is to create consistency, auditability, and a minimum standard across thousands of cases. They excel at saying what must happen. They are far worse at capturing everything a strong analyst actually notices: which sources they trust, when they widen the search, when a document feels off, when an explanation technically fits but still does not feel earned.
Procedures also carry the shape of the old operating model: fragmented systems, time pressure, queue pressure, and the hard limit of how much one human analyst can read, cross-reference, and hold in working memory at once. That is not a flaw in the procedure. It is how you design a process for humans.
AI changes the constraint set. Reading, searching, comparing documents, and tracing inconsistencies no longer have to be treated as scarce analyst time. Done carefully, with proper controls and human review, models can explore more context, test more hypotheses, and surface more inconsistencies than any single analyst could reasonably do case by case.
So if you simply automate the procedure exactly as written, you may gain efficiency. You will not unlock the full value of AI. You will just make the old bottleneck run faster.
The better question is not “Can AI follow the analyst playbook?”
It is: once the cost of reading, cross-referencing, and testing hypotheses collapses, what should the investigation become?
A second tempting approach: feed it historical Suspicious Activity Reports (SARs) and let it learn from outcomes. This breaks down too. You rarely have the full state of what the analyst actually saw during the investigation. A case that looks straightforward today might only look that way because information surfaced later. A fraud indictment that didn't exist when the original analyst made the call, news articles that hadn't been published yet. Hindsight can contaminate your training data. Also, regulators themselves acknowledge that SAR decisions can be subjective.
The architecture has four layers. The first is data: continuously enhancing the coverage, quality, and architecture of the signals the system depends on. The second is classical machine learning models that cluster and classify alerts to determine what type of investigation needs to run. The third is the investigation agent itself: a multi-agent system that orchestrates specialized agents to execute the investigation end to end. The fourth is a safety filter that runs independently of typology, ensuring no risk vector is missed regardless of how the alert is classified. Each layer is independently auditable and learns from the feedback provided by human reviewers.
Inside the investigation agent, specialized sub-agents run across the full case surface: alert context, customer and identity signals, access patterns, risk indicators, transaction behavior, source-of-funds, onchain activity, and public adverse media. Each writes its findings into a shared case memory. A coordinator agent reconciles and challenges them. When sub-agents disagree, such as when source-of-funds marks activity as “explained” while adverse media surfaces a recent indictment, the coordinator attempts to resolve these disagreements knowing the common patterns. The narrative agent prepares the final report with all collected evidence and suggested resolution. The last self-validation agent acts as a guardrail: if the system cannot support its conclusion with sufficient confidence or data quality, the case is routed to manual investigation instead of being surfaced as an automated result.
Before any of this touched a real customer case, we built what we call a “Golden Set” - historical cases with known right answers. "Known right answers" in compliance is harder than it sounds. It meant re-investigating old cases, getting multiple senior analysts to independently agree on what the right call would have been, then debating the disagreements until consensus. Months of work before we could even start measuring.
Here's an important part (for now) - cases currently get BOTH the AI's full investigation AND a senior human review. We didn't reduce scrutiny, in fact, we added more of it until it no longer proves valuable. Cases resolve significantly faster AND get more eyes than they ever did before. Every human correction feeds back into the model as a training signal. It gets better because it's wrong in front of people who know how to fix it.
None of this would have shipped without clearing structural blockers most financial institutions are still stuck on. Security and privacy sign-off to send customer data to LLMs at all. Senior compliance officer alignment on AI-assisted human decision making. Model Governance team embedded since December - they observed the entire Golden-Set Evaluation process and are running a formal validation review with our Internal Audit team now.
Today this handles roughly 55% of our US fraud case volume with significantly less analyst time per case. Time freed goes to the harder cases AI can't yet handle - and to teaching it.
Our internal compliance and quality teams are the ones who are building this system with the engineers, training it, validating it, and continuing to shape how it improves. In the process, they've developed skills that are incredibly valuable: how to design evals, how to think about model bias, how to think about human bias, how to architect human-in-the-loop systems, skills that are becoming among the most valuable at any company.
This entire project started ~6 months ago with a whiteboarding session between @galpa42 and I, and was built by an AI-pilled cross-functional and it’s just the first pod - there's a multi-month roadmap,rebuilding compliance from the ground up with AI. Huge thanks to everyone involved and congratulations to @galpa42 for shipping two babies to production this month :)
The future of high-stakes work is not AI replacing judgment. It is AI making judgment scalable, auditable, and continuously improvable.
Min-Si Wang retweeted
Introducing Privacy Cash SDK for Base, audited by @zigtur and @hashcloak.
Integrate privacy into your product with just a few lines of code, for private transfers and bridging.
Start building now.
github.com/Privacy-Cash/priv…
Min-Si Wang retweeted
For a more detailed explanation, take a look at the design doc: github.com/privacy-ethereum/…
Implementation: github.com/privacy-ethereum/…
Min-Si Wang retweeted
> "Our data is encrypted at rest and in transit"
> "We are SOC2 compliant"
> "We are based in the EU"
As we all know, none of these things matter when your data actually gets breached.
Modern compute infrastructure encrypts data at rest and in transit IF they are compliant
BUT as soon as the computation starts your data sits in plaintext inside a cloud you don’t fully control



































