0xApple (@0xapple_) header
0xApple (@0xapple_) profile picture

Independent Security Researcher at @sherlockdefi | Prev @kometverse @tradehotstuff

Joined December 2024
Photo by 0xApple (@0xapple_)Photo by 0xApple (@0xapple_)Photo by 0xApple (@0xapple_)Photo by 0xApple (@0xapple_)Photo by 0xApple (@0xapple_)Photo by 0xApple (@0xapple_)Photo by 0xApple (@0xapple_)Photo by 0xApple (@0xapple_)Photo by 0xApple (@0xapple_)
I audit smart contracts for a living 🔐 here's what that looks like in numbers: 💀 2 criticals found 🔥 11 highs found ⚠️ 23 mediums found 📋 57 total findings across 3 private audits + 14 public contests every finding = real money saved from hacks this is the most underrated skill in web3 🧵
5
1
135
4,641
I hate that security people are the first ones letting AI replace them. When AI agents and these advanced tools started showing up, we thought security was too critical for any agent to take over. What actually happened is that Web3 security people are so smart and adaptable that we turned it into a race to replace ourselves. Everyone is competing to automate everything and offer cheaper audits, and security itself became the last priority. I keep seeing people say their AI agent can do it better. You say that because you want to sell your product and your runs. Some of you even publish pricing tiers to cut the protocol’s cost. That trains clients to pay less every time and go looking for the next cheaper offer. We are diluting ourselves, and I really hate it. We should all be doing valuable work and asking for a fair price, so we can keep doing it properly and bring smart people into this industry. Low prices will only bring low-quality people in.
18
14
134
5,883
It’s always a pleasure to work with you guys 🤝
Audit three, is progressing well. Our superstar auditors at @PashovAuditGrp are logging findings as they go and we're tracking every one live. Initial report shouldn’t be too far off! On schedule. Chapter 2 goes live when it's done! Very excited.
1
16
1,040
🤝
Thankyou and your team as always!
1
69
Top read of the week
ChatGPT was really a big outlier in tech history. Imagine an alternative world where LLMs were developed with the exact same capabilities -- but nobody ever made a chatbot out of it. Instead Google integrated it with Search so it can give better results and summaries. You can do follow up queries that refine results of the initial query, including the summaries, and it also does some computations automatically. People hardly notice it. Github added a feature that you can enter a description of the code you want and it will find multiple OSS projects, fork them (keeping the license), recombine or integrating their code and even translate to other programming languages -- and give you something they call "initial version" that works well in many cases. They also add automatic suggestions to PRs, and suggested PRs for fixing/implementing open issues. They call it something like "advanced templating". It has mixed reactions among programmers but most organizations are using it to some extent. OSS developers actually like it. Wolfram Alpha now takes natural language descriptions of theorems and can prove or disprove many of them. Some people used it to find proofs for open problems. Some mathematicians worry the new generation is getting too dependent on it while the system doesn't really work every time. None even thinks about it replacing them. All the same capabilities, blended seamlessly into previously existing products. Nobody ever chats with an LLM. Nobody calls it AI. They don't have cute names -- they don't have separate names at all. They "live" inside boxes to perform tasks. They are components. This would likely be a much more productive world (I would switch to that world any time). I doubt anybody would be talking about exponential intelligence or worried about all jobs being replaced, much less about some doomsday scenario. If someone suggested these components were conscious people would laugh. It would be like saying like Google Search is conscious. Worse, like Big Tables or ranking algorithms are conscious. ChatGPT didn't ruin AI for people who tried it and never came back. It started training early adopters to think of it as "someone" instead of "something". The other companies then went even further. The entire concept of what LLMs are, what they are expected to be able to do or not, how reliable they are, what is their function in the workplace -- everything about them -- is built on top of the impression that they are basically like a person, because they can produce plausibly human-like interactions. A deception. If you want to use the LLM you're almost always required to interact with them by LARPing -- pretending it is "someone" and not "something", until you start using human words to refer to it, and forget it was just LARPing. That was a choice. Centering the development of LLMs as something that can pretend to be human instead of doing something useful. Building automated NPCs and pretending to be building God wasn't a given of the transformer architecture -- it was their deliberate choice. I think that was a bad idea. But it sure helped them raise money.
4
436
Reached out to a bunch of audit firms lately 🚀 Same story everywhere -- fewer audits, current team is enough to cover demand 📉 Why? Bear market dried up new protocols. The pipeline just shrank. What this means for independent researchers: the contest/freelance path gets more competitive, not less ⚔️ What protocols should do: this is actually the time to audit. Less demand = more senior eyes on your code, faster timelines, better rates 👀 The web3 security industry is at an inflection point.
3
23
2,338
Web3 has lost its ability to excite itself 🚨 Last four years of hype cycles make it obvious. 2023 | Account Abstraction 👻 ERC-4337 landed and the community was convinced that invisible wallets, sponsored gas and session keys were the missing piece for onboarding normies. 2024 | Intents and ZK 🔮 @AcrossProtocol and @CoWSwap shipping solver networks in production. ZK-EVMs were finally live: Scroll, Polygon zkEVM, zkSync all launched within months of each other and the rollup wars were at peak intensity. 2025 | Hooks 🪝 @Uniswap v4 opened up AMM logic to arbitrary customization and the DeFi community spent the year exploring what that design space could become. None of them delivered the inflection point the community expected. Each faded after its cycle 📉 2026 feels different 👀 There isn't an equivalent primitive. No breakthrough that Crypto Twitter has collectively decided will unlock mass adoption. That energy didn't disappear. It just moved to AI, every major release immediately shows people something new it can do. Meanwhile, Web3 security researchers are more excited about the latest AI model release than anything shipping onchain. Security firms are increasingly pitching AI as their next major differentiator 💸 The industry is borrowing AI's excitement because it isn't generating enough of its own. Web3 is probably fine long term. But it's in a very different chapter now 🫡
3
1
37
1,619
New researchers are scared of submission fees 🚨 The responses to my last tweet made that clear. I get it. But here's what should scare them more. Contest platforms are businesses. They're barely profitable, and AI slop is making the math worse. They will fix this. The only question is how. A fee is one answer. @0xPolygon Heimdall v2 contest tried another: personal invites to select SRs. Good experiment, though we don't know yet if it worked. But think about what each does to a newcomer. A fee is a tollgate, you pay and walk through. Invite-only is a locked door 🔒 If the alternative to a fee is being gated out entirely, a fee starts looking like the friendlier option ✅
🚨 The DRE App contest on @sherlockdefi had 1,973 nSLOC — and got 2,000+ submissions. Let that sink in. More submissions than lines of code 🤯 Preliminary judging at 100 issues/day = 20 days of work. For one contest. Sherlock is the last audit contest platform standing… and it's drowning in AI slop 🤖🗑️ Seems like $10/submission fee is the only way to fix this.
4
12
3,106
Web3 security firms are getting carried away with their AI claims 🧵 Firms are falling over each other to announce AI breakthroughs — some even going as far as saying their agent outperforms their own researchers. Bold claim. Yet not a single one has released their model for public testing. Funny how that works. What makes it stranger — these same founders charging $130k for a 14-day audit are now offering free AI pre-audits. Suddenly so generous 🤔 Here's the thing though, none of these companies have poached a senior AI researcher from Anthropic or OpenAI. These are security researchers who've fine-tuned frontier models on mostly public audit data. Respectable work, but let's not overstate what it is, because everyone is building on the same foundation: → Same frontier models (Claude, GPT-4o, Gemini) → Same public vulnerability datasets → No proprietary data moat → No dedicated AI infrastructure advantage So if everyone has the same ingredients, why are some founders acting like they've invented fire? The gap between any two firms here is a few weeks at best. And the context matters here — the web3 security industry is under real pressure right now. Firms that used to rely on a pool of contract researchers to handle overflow are now getting by with just their core staff. Clients have caught on, rates are being pushed down, and firms are quietly doing more with less. So when you see these AI announcements, ask the one question nobody wants to answer — where are the benchmarks? Where is the model? If it's already free, why not release it? Until then, please go easy on your AI 🫡
8
1
47
2,938
🚨 The DRE App contest on @sherlockdefi had 1,973 nSLOC — and got 2,000+ submissions. Let that sink in. More submissions than lines of code 🤯 Preliminary judging at 100 issues/day = 20 days of work. For one contest. Sherlock is the last audit contest platform standing… and it's drowning in AI slop 🤖🗑️ Seems like $10/submission fee is the only way to fix this.
9
1
31
5,775
Today I went rogue on bug hunting Spent 9 hours on ~100 lines of code ⏳ 4 of those hours on a single 30-line function First glance? Oh there has to be a bug here 🥳 I threw everything at it, hunted every edge case like my life depended on it Thought I'd found a bug, twice Each time, a tiny check buried in another function had already caught it. The audacity lol Couldn't break it. Couldn't even scratch it. Whoever wrote this, I see you 🫡
4
53
2,124
🚨 Here is some auditing alpha Don't blindly start reading code during a smart contract audit Spend the 1st hour on: 🗂️ State variables → what the protocol tracks 📡 Events → the actions it cares about ⚠️ Custom errors → edge cases devs already anticipated Only then does the code start making sense — this gives you the context required to actually start breaking it Most auditors skip this Most auditors miss things 🎯
3
5
48
1,570
I just ran Fable 5 and found a critical in Fable 5 itself
1
8
589
the energy i bring to the team
2
1
9
879
🧵 How to use AI in smart contract audits when your bar is already high: 1/ Compress protocol comprehension 🔍 Ask AI to explain a single execution path (liquidation, withdrawal, settlement), then branch into adjacent flows. You're not outsourcing analysis—you're accelerating time-to-model. 2/ Generate fuzz cases around invariants 👀 Give it an interface and a property you believe should always hold. Let it produce adversarial Foundry/Echidna tests while you focus on proving exploitability. 3/ Walk every if/else path with numbers 🔀 When logic forks, ask AI to trace each branch using concrete inputs and state transitions. Many accounting and edge-case bugs become obvious once you stop thinking in variables and start thinking in actual values. 4/ Use it for algebraic proof 📐 Before escalating a finding, ask AI to derive the constraints under which the exploit state is reachable. A few minutes of math can save hours chasing an impossible attack path. 5/ Turn findings into reports 📝 Once you have root cause, impact, and attack path, AI is excellent at converting rough notes into a structured writeup. Keep the judgment; delegate the formatting. ⚠️ AI shouldn't find bugs for you. It should shrink the distance between: "I don't understand this yet" and "I'm ready to break it" 🚀
1
7
53
2,322
I literally used #2 and #4 today. I understood the rough idea of a rounding issue, but wasn't sure whether it was exploitable. I asked AI to derive the constraints under which the rounding gap exists. A few iterations later, it produced a fuzz case that hit the exact edge condition and exposed a wrongful revert. AI didn't find the bug. It shortened the path from intuition to proof.
2
220
Every industry is slapping "AI" on its products and calling it innovation. Web3 security is no exception. The lowest-effort products are little more than an LLM with a Solidity-specific prompt and retrieval over past reports. The stronger systems combine LLMs with static analysis, symbolic execution, invariant generation, formal verification, exploit datasets, and years of audit knowledge. But there's an important limitation 🚨 These systems are fundamentally strongest when a vulnerability resembles something seen before, either directly or by analogy. The highest-impact bugs often emerge from novel protocol designs, unexpected interactions between components, or assumptions nobody realized they were making. Those are difficult not because they're hidden in the code, but because the security model itself hasn't been fully understood yet. AI is becoming an exceptional force multiplier for auditors. I'm less convinced it's close to replacing the people who discover entirely new classes of vulnerabilities.
1
2
10
541
The scariest files in any codebase aren't the complex ones. They're the simple ones nobody bothers to check 👀 Was auditing a smart contract so simple it almost got a free pass. I've seen this trap before. Simple code doesn't get audited. It gets assumed. Every auditor thinks the last one already checked. Nobody did. So I stopped. Looked closer. Ten minutes later — confirmed bug 🚀 This is the thing nobody tells you about security work: Complex code gets scrutinized. Simple code gets trust it hasn't earned. The bug isn't hiding. It's sitting right there, waiting for the auditor who actually stops. Next time you're about to skip a file — that's your cue.
1
11
529
Most auditors fail because they hunt for bugs 🚨 Stop looking for bugs Seriously Stop hunting Instead, obsess over: → Edge cases 🔍 → Accounting logic 🧮 → State transitions ⚙️ → Access control boundaries 🔐 Understand the system deeply enough, and the bugs will expose themselves to you 🎯 The best auditors don't find bugs They create the conditions for bugs to appear 🧠
1
5
51
1,367