Interested in software correctness. Cryptocurrency security researcher - https://xitter.gq/t.co/eZHZozB05V - https://xitter.gq/t.co/bLGkkx5E0e
0.0.0.0
Joined July 2022
- Tweets2.2K
- Following587
- Followers4.3K
- Likes2.1K
Pinned Tweet
I truly believe that we *can* make protocols rock solid.
It's not impossible. It's just a lot harder than people think it is.
Make them boring.
Don't ever upgrade without proving the whole system from scratch again.
Have a prover that you can run as part of development. Nothing gets released without all statements being proved.
This will slow development down to the speed of mollases, nay, tar pitch!
But this is actually what you want.
A formal methods *first* approach is probably what's needed.
I know people with these skills exist out there.
I know one project that took 25-man-years to prove just 8,500 lines of C was functionally correct. This sounds insane to most devs, especially in the age of AI assisted development.
But given what's at stake, this is probably what's required.
100proof.org/open-letter-on-…
100proof.org retweeted
I don’t want to say what the actual worst day in hacking history will look like.
But for sure the solution out of it will be based on social consensus, not technology.
100proof.org retweeted
Why software take so much time to start?
Why doesn’t it have an image in disk with everything it needs and just loads it in memory when you start the app?
I actually made that for some stuff back in the 90s made my own memory dumper and loader. And stuff would basically load instantly.
That was DOS though, of course a real OS would get in the way.
But way aren’t OSes built for that?
Why would you need to recreate from scratch all objects in a complex memory allocation?
We have so much memory now and ssd is so fast you could just keep everything in a persistent virtual memory forever.
We should be able to just have everything open at the same time and thousands of tabs open in the browser. And you basically never close or restart anything.
And if you run out of power when you turn on the computer again it just instantly loads the exact last state checkpoint.
Why arent we there yet with consumer OS?
Will the super intelligence please solve that?
100proof.org retweeted
Anything over 30 days with a critical bug is negligence. It now takes hours to weaponize an exploit.
Shut down the bounty. Blacklist the project across other bug bounty platforms. Publicly advise users to withdraw funds.
No more free bugs.
I also got a reply in 6 minutes.
Leaving a High severity issue unfixed for 200+ days,
collecting submission fees, then marking later reports as duplicates… that's not a bug bounty program. That's a trap.
If a project refuses to patch,
they should be forced to pay double the submission fee
to every researcher who finds the same issue later. A report escalated 211 days ago is still open. Do these programs even deserve to run a bug bounty?
100proof.org retweeted
Almost five months since this tweet, but enough has changed that I think it is worth revisiting a few points
> "The real question is: do LLMs get another capability jump, or just steady iteration?”
The first observation is that models did improve, but so far it has looked more like steady iteration than a significant step change. We will see where we are with astra and anthropic’s equivalent, and whether that continues to be the trend
> "My current bet is that within a few months, tools like Codex or Claude Security will be great at finding blockchain issues.”
I think it is pretty clear today that foundation models, and their baseline harnesses, can find cool blockchain vulns on their own. And they are good at it
Open models are also becoming interesting. There is still a gap with the best closed models, but the gap is now small enough that they are worth paying attention to
Interestingly, I see fewer of what I would call "traditional" false positives from codex/claude. The noise comes more from incorrect assumptions. The model shows a real path, and the reasoning might be correct, but it is based on a wrong assumption (e.g. incorrect config, impact that does not matter, an action no one would do, etc). So improving the results might become less about "can the model reason about this code?" and more about "does the model have the right assumptions?". And we are moving from "adding intelligence" to "adding the right context"
Specialized harnesses still have an edge today, but it is unclear to me how durable that edge is. Something like /goal with the right prompt is already powerful
That said, I have been experimenting a bit with harness self-improvement / customization. It is still experimental, but I could see a world where every team has its own customized harness, built from past audit reports, bug bounties, known assumptions, previous findings, etc. That could become a direct feature from codex/claude
If you build your own product, being better than claude/codex is not enough. You need to be better by enough to justify the friction of adopting another product. Maintaining that margin at low cost while the baseline improves every few months is challenging. That said, if a team manages to keep a lead, there is a lot of opportunity for them
But I think we are also past being impressed by claims like "our AI made $X million in bug bounties". Finding bugs is, ironically, the easy part. Finding useful bugs at low cost, with low noise and little human intervention is the hard part. That was true before LLMs, and it is still true now
So if you see a benchmark that only focuses on bugs found, without tokens spent and number of findings generated, you won’t know much. Similarly, a company saying its tool generated millions in bug bounties without saying how much it spends on inference per months and human review might not be that impressive
On the manual review side, I have now integrated LLMs pretty deeply into my own audit process. Building an LLM-wiki of a codebase on the fly is great. But there are downsides, and you might end up creating blind spots by not building your own proper intuition. I have tried different prompting strategies to reduce that. For example, explaining my understanding of a component to the LLM and asking it to challenge me, instead of asking it to explain the component directly. But nothing is perfect in my experience yet, so I have not found the right balance
That said, even if there are downsides to that approach, I do think LLM-assisted manual review is becoming the norm. So if you think LLMs do not create blind spots during audits, you are probably missing bugs. But if you think the solution is not to use them, you will probably get outperformed
Another area I am starting to look at more is LLM-generated code itself. I would not be surprised if we see new classes of bugs emerge from it. Sometimes models make mistakes that no human would make. Maybe models will improve enough that this disappears. But if not, there might be some interesting new attacks to consider
These days, in almost all my discussions I get asked what I think about AI and the future of security, so I figured I should share it here
Short version: I try not to have a strong opinion yet. We are clearly in a transition phase, and outside of people working directly on foundation models, no one really has a solid view of where this is going
Over the past months, LLMs improved a lot. The releases at the end of 2025 were a real step change. In practice, most people I know (myself included) have barely written code in the past 2-3 months. For security, we went from "this is fun" to "this is actually useful"
Right now, the best mental model I have is that we effectively jumped from having no tooling to having an advanced static analyzer or fuzzer. A lot of bugs that used to take time to find can now be surfaced quickly
Does that mean security researchers disappear in 2 years? Based on today’s tech, I do not think so. There are a lot of bugs to be found. Some are found by humans, some by traditional techniques, and now some by LLMs. But it does not mean all bugs get found. If anything, history suggests there are always more bugs than anyone expects, and that gap does not go away easily
The real question is: do LLMs get another capability jump, or just steady iteration? There are reasonable arguments both ways. To be honest, I do not have enough understanding of how these models evolve to have a confident answer. And anyone giving a very definite answer is probably overconfident, unless they are working directly on the models
Depending on that, the role of security researchers could change a lot, including the way we work. The demand could decrease if models get very strong at finding bugs. But it could also increase if the amount of code grows faster than the models’ ability to reason about it. We could even end up with a shortage of experienced researchers in a few years if fewer juniors enter the field while seniors move elsewhere. It is hard to predict because everything depends on how model capabilities evolve
On the business side, I am skeptical about "AI audit as a service". If models keep improving, it is hard to see how these companies compete with native offerings from OpenAI or Anthropic. Especially if those providers stop exposing raw capabilities and push everyone into their own products. I tried codex security, and while it is not perfect, it is clear where this is going. Mythos / Capybara seem to be around the corner, and it will be interesting to see how far it goes
My current bet is that within a few months, tools like codex or claude security will be great at finding blockchain issues, and they will integrate directly into most dev pipelines. At that point, the marginal value of an extra "AI audit SaaS" becomes limited
So what to do as a security researcher? Be adaptive. This is a transition period, and things will likely move fast in 2026. Stay curious, and keep working on skills that give you an edge. Regularly reassess where you are strong or weak, and where AI helps you versus where it replaces you. If you like challenges, see AI as one that pushes you to improve
Also, be careful with what people call "cognitive debt" or "brain rot". I was skeptical at first, but I do see it now. The more I rely on LLMs during an audit, the more I lose part of the intuition that I normally build while going deep into code. That intuition is still critical to find complex bugs. I have not found the right balance yet, but it is something to watch
It probably makes sense to revisit your view on LLMs every 3-6 months. I have already been wrong a few times on this, and I am fine with that, as long as I don’t get locked into a fixed view
Finally, a lot of people focus on the downside for security researchers. But there are also upsides. I can explore codebases much faster, build custom tooling easily, and spend less time on boring tasks. Maybe it’s my last few years/months as a security researcher, maybe not. But at least LLMs let me have some fun before doomsday 😅
100proof.org retweeted
I really don't know what the bottleneck is.
The BBP platform has already triaged it and escalated it to the protocols for their review, and it just sits there for like 6 weeks.
This has got to be a joke. As a protocol, you hire the BBP triagers for a reason.
Why can't they do this:
Run an LLM over every submission to
a) confirm it
b) ask a human to answer Y/N to "Send an email to the researcher telling them this is valid?"
I mean, seriously. Did NONE of them think of this? And why the hell wouldn't you do something like this?
This is a world where blackhats are running wild with their built-for-the-kill SOTA LLMs and harnesses.
It's absolutely wild for me to see that in the SAME FUCKING WORLD, a fully reproducible vulnerability, with an end-to-end PoC demonstrating critical impact... can sit in the “Triaged” state for six weeks without even a preliminary technical judgment.
If you are responsible for a project handling billions in value, please, for the love of God, get a grip.
Wake up to reality.
100proof.org retweeted
This is a world where blackhats are running wild with their built-for-the-kill SOTA LLMs and harnesses.
It's absolutely wild for me to see that in the SAME FUCKING WORLD, a fully reproducible vulnerability, with an end-to-end PoC demonstrating critical impact... can sit in the “Triaged” state for six weeks without even a preliminary technical judgment.
If you are responsible for a project handling billions in value, please, for the love of God, get a grip.
Wake up to reality.
100proof.org retweeted
I’m hearing elevated levels of dissatisfaction with submission fees, long triage times, and high levels of duplicates on bug bounty programs.
How about automatic full disclosure of bugs in 30 days or less?
Bug hunters shouldn’t have to waste time hunting dupes on deadbeat projects unable to protect themselves and their users expediently.
A reminder of how we got here and what will happen if legitimate security researchers are mistreated for too long:
100proof.org retweeted
I'm so freakin sick of how Claude talks.
It's not just annoying, it's infuriating.
The honest evaluation: it's the precise mechanism driving my current headache.
It's the load-bearing root cause of my daily stress.
It's genuinely painful, and that matters.
100proof.org retweeted
Bugs are found by reading code, right?
Reading is just the part you notice.
Ask Mike Tyson how to get good at boxing and he says "hit the bag more". Right, and completely useless.
"Just read the code" is the same answer. Reading is only what your top level notices. Underneath you are:
- forming a hypothesis and reading to DISPROVE it
- looking for what's missing, not what's there
- comparing what the dev meant vs what he wrote
- tracking one variable across every function that touches it
- mirroring paired functions side by side
- testing the boundaries
- reversing flows
- and a lot more
100proof.org retweeted
Interesting. But I find the whole "bitter lesson" thing quite nonsensical (So the article that expands on that idea doesn't make a lot of sense to me).
Bitter lesson -- the idea all we need to make computers do more things is more compute, and human ingenuity or concepts are mostly useless or even damaging.
Except that is obviously not true. We have all the software in the world, with exception of a subdomain conventionally called "AI" -- in which humans deliberately inputting human knowledge, concepts, abstractions, produces amazing results that are both reliable and scalable to a degree nothing produced by the "AI" field ever was.
In a sense you can define modern "AI" as "trying to make computers do things we don't know the algorithm for".
Which necessarily collapses into brute forcing. And the entire field of ML is basically about extracting a tiny bit more out of brute forcing, based on human assumptions, and human-curated or even human-generated data.
Take chess or go. We input into the machine all we *actually know for sure* about the game -- which is basically the rules. We then add to it the concept of grading positions based on future likelihood of success. *Then* the residue we find out works better if we just hit it with ungodly amounts of compute rather than try to guess heuristics.
Imagine if we did the same for a problem not classified as "AI" -- for example **sorting**. We input the concept of ordering. And just let the computer figure out what is the best way of ordering a bunch of random numbers. Train a massive neural network on samples of sorted sequences. Good luck with that. Instead we thought about it and came up with an algorithm that actually works.
Chess and sorting aren't fundamentally different as problems. Or chess and tic tac toe.
It could have turned out that chess is like tic tac toe, and there is an algorithm for perfect play -- and a human could even learn it and never lose to a computer (or other human), no matter how much compute you throw at it. It's an accident that chess doesn't have a small perfect play solution (though we might still find a longer one).
We don't know how to play chess correctly. The computer also doesn't know but it doesn't know faster. If some entity actually knew how to play chess correctly, no amount of compute in the universe would beat that good player.
So the real "bitter lesson" is that trying to make computers do things without knowing the algorithm is fundamentally intractable -- and giving it hints doesn't change the fact.
If you find out the algorithm it becomes suddenly tractable. If you don't, using human made heuristics hits a ceiling pretty fast.
You have to reduce the search space/domain to something that fits your brute-forcing budget, and you either choose how to do it yourself, or you let something randomly mutilate the domain for you (which is what LLMs do).
Then you are left with brute-force, and you can't actually escape the exponential cost of it, or the fact you can only achieve rudimentary approximations of an actual solution.
Your only hope is burning billions of dollars and praying it will be enough to approximate the solution sufficiently for your use case.
Maybe you will be lucky and you'll get something extremely useful. Maybe you will burn all the resources for nothing. All we know for sure is you won't come even close to what you would if you actually knew the algorithm for the problem.
That's the real "bitter lesson".
The only thing that works reliably is increasing compute -- because you don't know what you're doing.
If you did it wouldn't be "AI".
I don't think LLMs are a sort of "ghostly intelligence".
There is no such thing.
"Intelligence" means human intelligence or maybe animal intelligence. If you have to invent a new category to call LLMs intelligent, then what you're really saying is they aren't intelligent. The author of the article is just too emotionally attached to the thingy to say "LLMs aren't intelligent at all".
LLMs are an attempt to brute-force reproducing a sub-set of human behaviors without knowing the algorithm for the behaviors -- in fact without knowing anything about how humans do it.
And no amount of guesswork can turn that into a tractable problem.
Whatever you want to say about web3... it's not dull.
This is in my Top 10 Most Interesting IDMs I've Ever Read 🤣
etherscan.io/tx/0x6235601073…
100proof.org retweeted
I also have another theory about duplicates: some projects and platforms abusing nonsense slop submissions.
You disclose an e2e-proven exploit, but it gets marked as a duplicate because of the "root cause".
The slop report contains the vulnerable lines but no actual proof or has invalid claims.
With enough slop, you cover all the lines where a reasonable bug could exist. Then the project reopens the invalid slop submission, pays it as Low, and avoids paying the actual Critical.
That’s my worst nightmare. That shouldn’t happen ever.
Things could be worse.
LLMs are still so gormless that they will happily talk about their competitors accurately.
Woe betide us when this is no longer true.
100proof.org retweeted
Hate to break it to you, but the first LLM was created by Andrey Markov in 1913.
he tallied up 20,000 letters from a famous novel and computed
p(vowel | vowel)
p(consonant | vowel)
p(vowel | consonant)
p(consonant | consonant)
basically 'training' a bigram by hand
Bro just reinvented "In Praise of Idleness"
harpers.org/archive/1932/10/…
I’m cashing out all of the productivity gains from AI by working less, and it’s really starting to change my life
I rarely work past noon, because there’s simply nothing left to do. I do wake up earlier though because I’m so excited to get to work
I’m exercising more than I ever have in my life. Every afternoon either weightlifting, running, hiking, or paddle before the kids get home
I get a massage every week, and sauna twice a week, spending hours there with friends each time. The level of self care I’m doing is ridiculous. I’m almost too relaxed
My wife and I have a date night every week, our marriage is better than ever, and we decided to have a third kid, a son due in June
I have so much free time I’m starting to have to make up projects. I’m helping my friend start a nonprofit to promote local innovation and sustainability in our small Mexican town. Using Claude code to do all the writing, planning, and build a website for it
I’m spending more time talking to friends and family on FaceTime than ever in my adult life. I’m helping my parents and siblings with their work, heath, finances, and random problems, often using AI
Our social life is more active than even my teens or twenties, with at least 2-3 parties, dinners, or other gatherings each week. Everywhere I go in town I see people I know
We travel more often than ever, and take more vacation time than ever, though vacations are not as fun as the normal routine
The business is more profitable than ever, with a smaller team and less overhead. I’m able to pay my team better than I ever have. In no way is the business suffering
I say all this not to brag, but to show that there is another option for what to do with all the time and effort that AI frees up: you can pull back and live a more chill, social, connected life like humans were meant to
This is all due to AI, not because I got any smarter, wiser, or more productive. AI opens up new paths, but it’s still up to you to decide which one to take
A verifiable, public disclosure of an LLM-discovered vulnerability that would have put a truly eye-watering amount of funds at risk.
👏👏👏
Keep an eye on Apex from Cantina AI (@cantinaxyz)
We released a full report on the Batch amendment issue from last week.
The bug revolved around the signature validation logic of the Batch amendment.
It was caught before the amendment was activated by the autonomous AI agent Apex at @cantinaxyz.
Thank you to all validators and community members for the collaboration and swift response.
More details in the blog post here:
xrpl.org/blog/2026/vulnerabi…
