Beyond the Context Window: Why zkVM Auditing Still Needs a Human Map

Cloudflare's AI auditor, zkao, just dropped a major reality check on the "infinite context" hype. They identified a critical soundness bug (CVE-2026-46669) in the openvm-pairing library while experimenting on OpenVM's zkVM. The vulnerability is serious: it allows a malicious prover to forge any pairing equality. But the real story isn't just the bug itself—it’s the technical wall the AI hit while trying to find it.

The Context Window is Not a Magic Bullet

For anyone trying to automate security audits, this is a necessary "stop and think" moment. We’ve been sold the idea that if we just give an LLM a bigger context window, it will eventually "understand" any codebase. This research proves that’s a fantasy for high-complexity systems. The team found that naive LLM setups struggled with the dense, interlocking dependencies of zkVMs compared to simpler cryptographic libraries. Even with 300K or even 1M tokens of context, the models couldn't navigate the codebase effectively enough to surface the bug on their own.

The takeaway? For high-stakes infrastructure like zero-knowledge virtual machines, "more data" is not a substitute for "architectural awareness." If the model doesn't understand the underlying logic of how the components interact, it’s just staring at a mountain of text without a compass.

Context Engineering: Building the Roadmap

So, how did zkao actually crack it? They didn't just throw more tokens at the problem; they used context engineering and encoded expert methods. Think of this as providing the AI with a specialized roadmap rather than just giving it the keys to a library. By structuring the context to mirror how an expert would navigate the code, they were able to guide the AI through the complexity of the openvm-pairing library.

Crucially, the AI didn't just spit out a finished report. It produced a "candidate finding." This is a vital distinction for builders to understand. That candidate finding still required human experts to step in, validate the exploitability, and confirm the actual impact. It took nine and a half hours of scanning to get there, proving that while AI can handle the heavy lifting of wide-scale code traversal, it still needs a human-designed framework to stay on the rails.

The "Augmented Human" Reality Check

We need to stop talking about autonomous AI security and start talking about augmented human security. If your AI can only produce a "candidate finding" rather than a verified report, your security posture still rests entirely on the humans who have the expertise to vet those leads.

The real story here is the bottleneck of validation. The practical utility of these tools right now is as a high-speed filter for human experts—a way to sift through the haystack to find the needles. The interesting question isn't whether an LLM can find a bug; it's whether it can find the right bugs without being led by hand through a maze of complex dependencies. If the human expert is still the ultimate bottleneck, the impact of these tools at scale is limited by how many high-quality leads a human can realistically verify in a day. We aren't replacing the auditor; we're giving them a much faster shovel.