Many people have claimed that with AI-assisted bug finding, secure code (and hence trustless anything) will be impossible.
I have a much more optimistic take, and AI-assisted formal verification is a major part of the reason why:
https://t.co/0ceMBZ6uqj
Many people have claimed that with AI-assisted bug finding, secure code (and hence trustless anything) will be impossible.
I have a much more optimistic take, and AI-assisted formal verification is a major part of the reason why:
https://t.co/0ceMBZ6uqj
CLAP lets you write a ZK circuit directly in the Lean proof assistant. The compiler lowers it to a constraint system, with soundness and completeness guaranteed by construction.
Properties proved at the source level carry through to R1CS.
This is a great illustration of why "generative AI finding security bugs everywhere" is a non‑equilibrium phenomenon. Today, there's lots of software out there that wasn't built under agentic scrutiny.
Going forward, pretty much all software, including smart contracts, will be subjected to increasingly sophisticated and thorough agent audits before deployment – which will hopefully squash these types of bugs ahead of time.
(h/t @eddylazzarin)
I think my point may have gotten lost. The takeaway was not that agents can now do all this automatically. It was that they still fall short.
Yes, most of the cases are pre-cutoff, so contamination is possible. But even if that gave the agents some advantage, they still failed to produce PoCs in more complex cases.
My point was that people sometimes talk as if AI can already automate the whole process. But if it can spot a vulnerability and still fail to produce a PoC, human review is still needed, which means the model is assisting experts, not replacing them.
@jack__sanford@a16zcrypto yes, since this experiment uses historical data, there's a chance of knowledge contamination. so failure rate is the only metric i'd really trust here.
@jack__sanford@a16zcrypto vuln discoveray rate was higher without hints (i don't have the exact number rn). will update once i rerun it with the newer model.
We've been exploring whether AI security agents can replace human experts.
Our takeaway so far: not yet.
For price manipulation, one of the most complex DeFi attacks, generating exploit PoCs is still a bottleneck.
Would love your thoughts.
1/ Moltbook is a great experiment, but it brings up a lingering question: How do we secure vibe coded applications? The answer, hopefully, is to have some AI help secure code written by AI, but the specific details are still an active question.
Recent Balancer + yETH exploits share a common pattern: tiny numerical edge cases becoming real vulnerabilities.
We outline the mechanics and argue for a missing defense: runtime checks on precision + invariants.
This needs to be a first-class design requirement in DeFi.👇
@AlexQuellsIt Most auditors can identify core invariants, but they rarely push for runtime enforcement unless the client asks. So I recommend asking them to include those invariants explicitly in the audit report; it changes the outcome.
@ittaia Love that point! Time delays could be the safest and most robust route when composability isn’t critical. UI would need some thoughtful design, but totally worth it. Really appreciate the feedback!
8/ As DeFi math gets more complex, this line of defense becomes essential.
Even if an unforeseen numerical subtlety slips through, the protocol should never enter a state that violates its economic invariants.