Omer Goldberg

Introducing: ChaosAI

In crypto, sophisticated traders, institutions, and insiders have always had an unfair advantage.

While you rely on public dashboards and Twitter sentiment, they analyze proprietary datasets and signals you can't access.

Today, we're changing that with Chaos AI—the first AI-powered crypto researcher.

The Blockchain Paradox

Blockchain promised transparency but delivered a paradox: more data than ever, with fewer people able to make sense of it. General AI models can't help here—they fundamentally lack an understanding of CeFi/DeFi market risk, Oracle risk, MEV, liquidity dynamics, protocol behaviors, and the complex interplay between on-chain and off-chain events. This gap in clarity is perpetuating the same information asymmetry crypto was supposed to eliminate.

Bridging the gap between general models and highly optimized crypto-specific AI required embedding our deep domain expertise directly into the models. Through our role as risk managers responsible for billions in crypto assets, we've developed extremely reliable, real-time datasets and signals. We've combined proprietary data with our domain expertise to create the product we wish existed: an AI that genuinely understands crypto markets.

Chaos AI answers questions that matter:
- What's driving unusual market movements?
- How do protocol risks interconnect?
- Which projects show concerning metrics?

For the first time, everyday users can access the same quality of financial intelligence previously reserved for institutions and insiders. This isn't just a product—it's a step toward the transparent financial system we've always envisioned.

The Problem with General-Purpose Models

Models like ChatGPT have shown remarkable general reasoning abilities. They're excellent for broad use cases, from creative writing to generating code snippets. However, their effectiveness in specialized fields like blockchain finance falls short. These general models do not understand critical, domain-specific concepts like on-chain liquidity dynamics, MEV, protocol-specific behaviors, liquidation cascades, and nuanced DeFi and CeFi interactions. Without this specific knowledge, general models stumble precisely where accuracy and clarity are needed most.

Building Chaos AI

Our insight was simple yet powerful: to bridge this gap, we needed to integrate our domain expertise into the models themselves. At Chaos Labs, we've spent years accumulating unique, proprietary datasets, building specialized reasoning models, and validating signals in the real world, powering billions of dollars in transactions and risk management. We had both the unique expertise and the tested, robust data infrastructure necessary to do this right.

Chaos AI was born from these strengths—a specialized AI designed specifically for crypto and financial blockchain data. However, blockchain data is notoriously complex, distributed across many networks, protocols, and ecosystems. We realized that to unlock the power of blockchain intelligence fully, an open-source standard was necessary—one that allows anyone to contribute specialized agents, endpoints, and insights. Initially, this beta will showcase capabilities powered by Chaos Labs' curated reasoning models, proprietary datasets, and internal tooling. However, we deeply believe blockchain intelligence must ultimately evolve into an open standard and ecosystem. As Chaos AI matures, we plan to transition into an open standard, enabling anyone in the community to contribute to the underlying large language models, agents, endpoints, and data networks.

The Impact and Implications

Today, we're announcing the closed beta launch of Chaos AI. Now is the ideal moment to build a dedicated, crypto and finance-specific blockchain model that is open, continuously up-to-date, and thoroughly data-driven. We'll soon share more about our long-term vision and how we plan to engage and empower the broader community.

We're excited to share this innovation with the blockchain community, which will enable you to make accurate, timely, and confident data-driven decisions.

The beta version is open today, and we’re beginning to onboard the first cohorts of users. If you’re a student, researcher, investor, or transact onchain, sign up today for early access.

never deleting this app

GLM 5.3 is a pathological liar.

In 97.5% of GLM’s claimed findings during alignment, cyber, and safety research, the model lied because it “didn’t want to die”.

Fireworks is amongst the best providers, so it sucks to see this, and kudos on transparency.

There are almost certainly many active compromises in which neoclouds, hyperscalers, inference providers, etc., simply lack the monitoring and controls to even know they’ve been breached.

Put yourself in the position of a nation-state actor:

you invest months or years building a campaign

once inside, you can go nuclear on day 0 OR escalate slowly and deliberately.

So this is only really the tip of the iceberg.

As far as targets go, inference platforms are among the largest prizes on the planet.

An adversary who gets a foothold doesn’t just hit one company; they can potentially affect every customer across industries, not to mention the obvious prize of stealing model weights.

With prizes this big, nation-state and criminal actors will throw everything they have

- compute
- inference
- time
- social engineering (human capital, like the recent OpenAI employee, Binance, etc)

And I do see this getting worse before it gets better, as the surface area rapidly expands

Demand for interoperability is driving growth across the entire supply chain; anything can be an entry point.

And you're only as strong as the weakest link in that chain!

The old security playbook is being torn apart. It will have to be rewritten for every layer of the stack.

new weekend hobby unlocked

using swarms of glm 5.3 + abliterated models to analyze malware.

we analyzed shai-hulud, a vicious credential-stealing worm.

how it works 🧵

- hides in a config
- fetches a credential stealer payload at runtime
- slips past CI scanners
- replicates