Will AI Take the Security Researcher's Job?

Your Security Career in the Age of AI

We are still in the trial-and-error phase of AI and Web3 security, and nobody has run it long enough to know how it settles. People are already asking the obvious question, and the ones who do this for a living ask it most anxiously. Does AI take the security researcher’s job? The honest answer depends on which layer of the system you mean. On the contract layer, where someone once read the code line by line, AI is taking over that slow manual work, and the expert is moving up the ladder to the work a machine cannot do.

Consider what the tools already do. They run as clusters of agents that watch a codebase around the clock, and formal verification is moving into the code-generation loop itself. Certora’s AI Composer checks AI-written code against mathematical rules before it runs, and Vitalik Buterin has argued that pairing AI-generated code with machine-checked proofs is the most credible path to secure software as AI grows better at breaking it. The floor under a newly deployed contract is rising, and the standard for calling one secure has risen with the tools that test it.

This is what moves the auditor up the ladder. The line-by-line reading, the checks against known patterns, the easy findings, all of it now goes to the agents, and what lands on the researcher’s desk is a pile of machine output that someone has to sort. The machine finds the obvious bug faster than any person can, so the skill that matters now is judgement, knowing which of the findings is real, which is noise, and which one matters for this particular protocol. Because both attacker and defender hold the same tools, the advantage belongs to whoever can read what the machine produced and understand it, and that judgement sits well above the grind it replaced.

Photo by Levon Vardanyan on Unsplash

But the contract / code was never the whole system, and the audit that vouches for it has a narrower reach than the badge suggests. An audit certifies a frozen commit, a snapshot of the code as it stood on the day it was read, while the protocol goes on shipping upgrades and new integrations until that snapshot no longer matches what is running. Formal verification has the same limit. Vitalik makes the point plainly in the same essay, that you can prove one part of a system correct and still lose to a bug in the parts you left unproven, or to the hardware underneath. The proof describes a map, and the attacker is loose in the territory the map leaves out.

That area is where the money now leaves. The largest losses of the past two years came from outside the contract code entirely. More than $3.4 billion was stolen in 2025, and the single biggest theft, roughly $1.5 billion from Bybit, came through a compromised third-party integration that tricked the people holding the keys into approving it. Closer to the shape of a normal protocol, one compromised laptop at Humanity Protocol carried enough keys to clear two multisigs, and the attacker used them to seize the bridge and upgrade it to a malicious version. No contract was broken in either case. The way in was a stolen key, the Shai-Hulud worm spreading through the npm registry, an employee who clicked the wrong link. Meanwhile the losses on the contract layer were drifting lower, because that is where the tooling and the incident response had improved.

We learned to defend the contract because that is where the whole field has trained its attention, and we left the rest of the stack, the keys and the pipelines and the people, to look after itself. That is why it stayed open. I set out the full case for this in a report on this year’s operational losses.

---

So moving up the ladder is only half of what the job now asks. The expert also has to move outward, to the off-chain code that nobody formally verifies, to the access controls sitting above the contract, to the deployment pipeline, and to the people who hold the keys. These are the failures that live above the code, and they are the ones that have been paying out. Used well, AI arms that expert, the way the scanner and the fuzzer armed the generation before, and the people who read it as a reason to do without the expert have the lesson backwards. They overestimate what the model does on its own and underestimate how much still rests on a person who understands where the real risk has gone. We are early, and the tools will only get better. My bet is on the people who treat the model as a tool, learn to aim it well, and point it at the parts of the stack the badge on the website never covered.

— Adrian

P.S. There is one thing the ladder model leaves unresolved. The grind we are handing to the machines was also how people learned the craft. You read ten thousand lines of someone else’s code by hand, and somewhere in that tedium you earned the judgement to know which finding matters. Hand all of it to the agents, and the junior never does the reps. I do not have a clean answer to where the next generation of experts comes from once the training ground is automated, and if you are running a team, that is the question I would be sitting with.