Source: The Defiant — https://thedefiant.io/news/hacks/openzeppelin-pushes-back-after-ex-cto-declares-all-of-defi-unsafe
The claim is bigger than the evidence we have#
OpenZeppelin is pushing back after its former CTO, Manuel Aráoz, warned that AI coding agents have made DeFi impossible to defend.
The firm did not deny the underlying risk. Its position, as reported by The Defiant, is narrower: AI-assisted coding is a real threat to smart contract security, but not an automatic death sentence for DeFi. That distinction matters. One version says the category is structurally unsafe. The other says the security model is under pressure and needs stronger controls.
The public tension is useful because it strips away the easy marketing language around AI development. Faster code generation does not mean faster understanding. In smart contracts, that gap is expensive. A bad web app can often be patched after deployment. A flawed DeFi contract can move funds before governance, auditors, or users have time to react.
Aráoz’s warning lands because DeFi already has a long record of bugs, exploits, broken assumptions, and incentive failures. AI coding agents add another failure mode: they can produce plausible code at scale, including code that looks clean enough to pass a shallow review. That does not prove all DeFi is unsafe. It does make weak review processes harder to excuse.
Why AI coding agents change the security equation#
The core issue is not that AI writes “bad code.” Humans do that too. The sharper issue is throughput.
AI tools can increase the amount of smart contract code being produced, forked, modified, and shipped. If review capacity does not rise with it, the industry gets more attack surface with the same bottleneck at the end. Audits, formal verification, internal review, and bug bounty programs all depend on time, context, and adversarial thinking. A coding agent does not remove that work. It can hide how much of it is now needed.
This is especially dangerous in DeFi because contracts rarely exist in isolation. A protocol can be affected by oracle behavior, governance privileges, token mechanics, bridge assumptions, liquidation design, external integrations, and upgrade paths. AI can help draft components, but security failures often live in the seams between components.
OpenZeppelin’s response appears to defend that distinction: the threat is real, but manageable. That is a more defensible position than blanket reassurance. “Manageable” still implies work. It implies slower deployment where risk is high, better testing, clearer threat models, and less faith in generated code simply because it compiles.
What not to overclaim#
The source does not establish that AI-generated smart contracts are currently causing a measurable wave of DeFi exploits. It also does not show that OpenZeppelin believes existing practices are enough without change.
It is safer to read this as a dispute over risk framing. Aráoz is making a categorical warning: AI coding agents have pushed DeFi beyond defensibility. OpenZeppelin is rejecting the categorical part while accepting the risk premise.
That difference is not cosmetic. If DeFi is “impossible to defend,” then audits, libraries, monitoring, formal methods, and cautious deployment are reduced to theater. If the threat is manageable, then the question becomes which teams are actually managing it and which are using AI to ship faster without increasing review depth.
The second question is more useful for users, investors, and builders. It can be tested against process. Does the team publish audits? Are privileged roles documented? Are upgrades timelocked? Is there an active bug bounty? Are dependencies maintained? Is generated code treated as untrusted until reviewed? Those signals are imperfect, but they are better than slogan-level confidence.
Practical checks for builders and users#
For builders, the safe default is simple: treat AI-generated contract code as hostile draft material, not as finished engineering. Require human review by someone who understands the protocol’s economic assumptions, not only Solidity syntax. Add tests for invariants, edge cases, and integration behavior. Keep generated code paths traceable so reviewers know what changed and why.
For users, the practical lesson is not to abandon every DeFi interaction. It is to price in development discipline. A protocol that ships rapidly with unclear audits, vague upgrade controls, and little disclosure deserves more skepticism in an AI-assisted coding environment. The same goes for forks that inherit old code, add new features, and present the result as battle-tested.
OpenZeppelin’s pushback is credible only if “manageable” remains a high bar. AI can make smart contract development faster. It can also make security debt arrive faster than teams can name it. That is the real pressure point.