Most market observers read the headline—'Wiz CEO builds investment empire in AI cybersecurity after Google deal'—and see a typical founder cashing out to play venture capitalist. But as a smart contract architect who has spent 18 years dissecting composability failures, I see something else: a signal that the current AI security stack is structurally broken at the abstraction layer.
Let me be specific. The deal itself is a black box: Assaf Rappaport, after walking away from the $23B Google acquisition, reportedly begins deploying capital into AI-native cybersecurity startups. The exact targets, check sizes, and technical theses remain undisclosed. But the very vagueness of the narrative—coupled with the fact that a crypto-native media outlet (Crypto Briefing) is the sole source—suggests we are looking at a carefully staged PR signal rather than an investment thesis. In blockchain terms, this is a governance proposal with no on-chain data to verify.
Context: The Composability Crisis in AI Security
To understand why this matters, you need to map the current AI security landscape onto DeFi’s composability pyramid. In DeFi, we learned the hard way that composing smart contracts without formal verification leads to catastrophic reentrancy and oracle manipulation. The same pattern is unfolding in AI: LLM guardrails, agentic frameworks, and verification layers are being stitched together like Uniswap pools, each with its own trust assumptions and attack surfaces.
Consider the typical AI security stack today: (1) prompt injection filters, (2) output constraints, (3) cryptographic integrity proofs (e.g., zkML), and (4) governance oracles. This is analogous to a lending protocol composing a flash loan aggregator, a liquidation engine, and an oracle. Each layer introduces latency, cost, and failure points. The Wiz CEO—a man who built his career on cloud security’s “single-pane-of-glass” philosophy—likely sees the same fragmentation.
Core: Hypothesis-Driven Simulation of the Investment Logic
Let’s run a simulation. Based on my audit experience with zero-knowledge proving systems, I can infer the following:
Assaf Rappaport’s investment criteria likely mirror the engineering-first pragmatism that made Wiz successful. Wiz’s core innovation was agentless scanning—removing the need for third-party software on customer endpoints. In AI security, the equivalent is removing the need for model-level access. The most scalable security startups will be those that operate at the verification layer, not the model layer.
Specifically, I believe Rappaport is betting on zkML (zero-knowledge machine learning) startups. The reason is simple: Composability isn’t a feature; it’s a ecosystem-wide invariant that must be enforced at the cryptographic base layer. Wiz’s success came from abstracting away cloud complexity; the next wave will abstract away inference trust. zkML allows a model to prove a computation was performed correctly without revealing the model weights or the input data—a perfect fit for regulated industries like healthcare and finance where AI decisions require audit trails.
But here’s where the simulation hits a constraint: current zkML provers are too slow for real-time inference. A 2025 benchmark showed that proving a single transformer inference (GPT-2 scale) takes 12 seconds on a 32-core machine. That's 12 seconds of latency in a world where a flash loan executes in 0.5 seconds. The trade-off between latency and trust is the same one we faced with ZK-rollups: sequencer centralization or slow finality.
Contrarian: The Blind Spot in the Investment Empire Thesis
The real risk—the one no CNBC headline will capture—is that the Wiz CEO is inadvertently funding the same centralized architectures that he escaped from at Google. Layer2 sequencers are basically single centralized nodes; “decentralized sequencing” has been a PowerPoint slide for two years. Similarly, many AI security startups I’ve audited rely on a single API key from OpenAI or Anthropic to provide “guardrails.” That’s not security; that’s a pass-through to a centralized model provider.
If Rappaport’s portfolio companies replicate this pattern, his “investment empire” will be built on sand. The market will eventually realize that an AI security startup that relies on a single inference API cannot survive an API outage, a pricing change, or a model censorship event. We don’t build anti-fragile systems by layering centralized dependencies. Post-ETF approval, BTC has become Wall Street’s toy; the original vision of peer-to-peer electronic cash is dead. The same can happen to AI security if we don’t enforce cryptographic sovereignty at every layer.
Takeaway: The Vulnerability Forecast
In the next 12–18 months, we will see a massive re-valuation of AI security startups based on their architectural decentralization. The ones that rely on closed-source model providers will be treated like centralized exchanges—prone to regulatory capture and single-point-of-failure attacks. The winners will be those that embrace verifiable computation and permissionless composability.
As an ecosystem, we need to ask ourselves: Are we building a Layer2 for AI security, or just another centralized sequencer? The answer will determine whether the next $23B exit is a technological leap or a regulatory arbitrage.