How Ethereum's Top Bug Hunter Finds Unique Bugs (w/ Alex Filippov)
文章目录
Alex Filippov breaks Ethereum for a living.
He won the Ethereum Foundation's $2M Lighthouse competition, taking home $203,733 with a single High-severity finding. He won the Ethereum Fusaka Upgrade contest with two High-severity bugs that nobody else found: one in Prysm and one in the KZG proof verification library c-kzg-4844 (both in the contest report). He is #3 on Cantina's all-time leaderboard with more than $542,000 in winnings, and #9 on the Ethereum Foundation's consensus-layer bug bounty leaderboard, the program that pays for bugs in the clients that keep Ethereum running. Before all of that, he earned a degree in cryptography.
Despite a very busy schedule, both at B612 Security and in his own bounty work, we got him to walk us through exactly how he works: where he starts, how much of the job he hands to AI, how he picks a program and when he walks away from one.
If you're hunting with generic prompts and shortcuts, his experience comes with a warning: you'll find bugs, but they'll be the same ones everyone else finds, and duplicates pay little or nothing. In his words: "Finding a valid issue is actually quite easy nowadays. Finding a non-duplicate one is hard :)"
How he starts an audit
I usually spend a couple of hours skimming through the scope, getting a high-level understanding of the protocol or changes, and splitting the work into chunks. I write a rough plan with what seems like a reasonable review order and then start auditing.
At that stage, I might read docs or code, watch videos, consult with AI, or use whatever else helps me build an initial understanding of the scope.
Private audits: every line by hand, then AI
My workflow differs quite a bit between private audits and bug bounties. At the moment, I'd say bug bounties are around 30% manual / 70% AI, while private audits are closer to 60% manual / 40% AI.
In private audits, I usually split the scope into parts first (modules, PRs, etc.) and go piece by piece. For each part, I first review every line manually and note any bugs or suspicious places. Then I do several passes with the AIs I use (Codex, Claude, GLM), investigate the resulting leads, and write draft reports. Closer to the end, I review everything again, clean up the reports, and submit them.
I haven't participated in contests for about a year, and that was mostly before I started using AI heavily, so I don't think my experience there is very relevant today.
Bug bounties: one impact at a time
The main difference between the two for me is coverage.
In private audits, I usually focus on maximizing coverage. I read the whole scope line by line and make sure I'm not missing anything. In bug bounties, that usually doesn't make much sense. It's often more efficient to focus on specific impacts or parts of the codebase that look promising.
So I focus on impacts and review the parts of the codebase that are relevant to each particular impact.
Before August, my approach was closer to private audits, but I focused on one impact at a time and didn't try to read every line. It was more like a stochastic search driven by my intuition and AI, focusing on where I felt bugs causing a particular impact could be. For example, for an RPC crash I'd focus on the RPC layer; for a chain split or halt, on consensus-critical parts. It was probably around 50/50 manual and AI.
The tool he built after a month of duplicates
That resulted in tons of duplicates in August, so in September I spent quite a bit of time building some tooling.
Now I have a tool that helps me with bug bounties. It's nothing fancy, just a UI + backend + DB that can run agents under the hood to generate and verify leads, build end-to-end PoCs with a running localnet, write reports, etc.
I wouldn't call it an autonomous audit agent, since I didn't embed audit skills or common bug patterns into it. It just automates the parts of the process that are reasonably automatable.
So now I mostly read the code myself, generate ideas or leads, and send them to the tool, which verifies them, expands on them, builds PoCs, writes reports, etc. Then I just triage final reports manually and submit them.
Where the non-duplicates come from
That said, I haven't had much success using it with generic patterns. It finds them quite well, but they mostly end up being duplicates. So I still think you either need to be faster than everyone else, which is probably not a race you can consistently win long term, or have genuinely valuable manual insights.
In my experience lately, almost every report I've had that wasn't a duplicate started from one of my own ideas. AI can and should explore, research, write PoCs, and prepare reports. But you still need to understand the codebase and bring some unique ideas if you want to consistently find non-dupes.
There is, of course, an alternative strategy where you use AI at scale and try to cover as many bug bounty programs as quickly as possible. This will probably still produce results even with fairly generic prompts and patterns, but I don't think it's the best long-term strategy. Eventually, this will likely just become the new baseline.
How he guides AI
In both audits and bug bounties, I try to have a deep understanding of the protocol, build a reasonably complete threat model, and guide AI accordingly.
As for fancy prompts or techniques, I haven't invested much time in that. I mostly just try to ask the right questions with the right context.
I prefer to give AI more freedom while staying quite involved in the loop myself, both in bug bounties and private audits.
I might be wrong, but I feel like deterministic approaches will, if they haven't already, become part of the baseline. There is still a lot to build there, but bigger teams are probably better positioned to do it successfully.
What I try to provide instead is a unique perspective on the protocol based on my experience and unique (I hope :)) ideas.
Picking a program and knowing when to stop
Most of my bug bounty wins came from two protocols that are quite similar, so I mostly stick to several programs that are interesting to me and where I feel the process is fair. I plan to review a couple of new projects soon, but I wouldn't say I have that much experience picking programs.
I usually stop when I'm running out of ideas and can't generate anything new for quite a while, like a week, or when I feel like I'm starting to burn out on that particular bounty.
Bug bounties or private audits?
I don't think there is a single answer that works for everyone. It really depends on the person.
I still think bug bounties can be the better choice for some people, but probably for fewer people than a year ago. They have become much more brutal. In my experience, the number of duplicates has increased something like tenfold.
So in general, private audits are probably the safer choice for most people.
For anyone who still wants to start hunting:
Beyond the obvious stuff like having enough runway for a few months and being prepared for it to be rough, I'd say use AI heavily, but don't rely on generic prompts and patterns.
When months go by without a valid finding
I think the best, although somewhat paradoxical, advice here is to stop hunting for a while and take a break.
It's very easy to get into a vicious loop where you can't find anything, so you decide you need to work more and harder, and then you still can't find anything. In my experience, taking a break is often the best thing you can do, but it can be very hard, especially when there is financial pressure.
Finding a valid issue is actually quite easy nowadays. Finding a non-duplicate one is hard :)
Having some financial security helps a lot. It also helps to focus on the process rather than the outcome. But I don't think there is any real way to avoid the uncertainty and disappointment. You have to live with it and keep going :)
Where web3 security is heading
I don't have much experience with security firms, and predicting anything nowadays is pretty hard, so I don't have a strong opinion on this.
I think the AI part will only grow, but I feel like humans will still be involved in generating ideas for AI to check, building AI harnesses, and verifying the output for quite some time.
What it takes to be a good auditor
I think you need to be curious, patient, and persistent.
Work with researchers like Alex
Alex's answers point to the same conclusion we built B612 Security on: AI finds a lot, but the bugs that matter most still start with a senior researcher who understands the system deeply and knows where to look.
That is the kind of researcher who reviews your code at B612 Security. If you are building or upgrading an Ethereum client, consensus or networking code, a ZK or cryptography-heavy protocol, or anything that relies on KZG proofs or data availability, reach out for a security review.



