AI is becoming awfully good at finding bugs, and cryptography is starting to reorganize itself around this fact.
Just watch the OG channel this week. Thursday (October 22), Bitcoin Core merged a security policy change for bugs “that can be found with a simple prompt to an LLM.” Instead of remaining embargoed for a year, these vectors are now repaired and shipped as soon as possible. Like the author of the change, Antoine Poinsot, put it onLLMs “dramatically lower the bar for finding certain categories of problems.”
Meanwhile, two critical PHP flaws discovered with the help of AI models remain ~1.3 million hosts exposed to the Internet vulnerable. And today, Initia arrested its entire network after an attacker exploited what co-founder Zon called a “day 0 bug” to create tokens and fill approximately $2.7 million. There is no confirmation yet that AI was used for this hack, but it seems safe to assume that it was. a few ability.
DefiLlama’s hack database, which I delved into for my data markets article this week, now records 112 hacks over $1 million so far in 2026, up from 70 in 2026. all of 2025. And Anthropic smart contract research found that revenue from simulated frontier model exploits will double approximately every 1.3 months in 2025.
So what to do when attackers have tireless and constantly improving bug hunters? Crypto is converging on two very different answers.
The first answer is to stay agile, that is, maintain the ability to patch or shut things down quickly.
Starknet is the clearest example of this philosophy today. Thursday, his community vote looking to become an independent post-quantum L1, in part so it can upgrade its cryptography on its own schedule. Like StarkWare put it on“we can’t count on Ethereum to act fast enough to keep us safe.” CEO Eli Ben-Sasson now stakes the arrival of the quantum at “2 years”, arguing that AI accelerates the timetable.
— (@)
Ethereum researchers are also on alert. Earlier this month, Justin Drake exhorted the industry to plan a “bunker mode”, i.e. moving funds to new, safer addresses, after a wave of AI-based mathematical breakthroughs.
But in all honesty, agility comes at a cost. Each admin key and upgrade path is Also something to attack or rely on. And changing the rules can create its own bugs, which is why Ethereum removed Glamsterdam from its Hoodi testnet this week following research showing its gas changes could alter the behavior of deployed contracts.
The second response is to narrow the target. Here’s how Ethereum developer Ross (z0r0z) presented it on Thursday:
Do you like this article?
Subscribe to Bankless or log in
— (@)
It’s a bit ambitious, of course. Yet the idea is fundamentally an expression of Ross’s exit testing philosophy, that is, can a system continue to operate safely even if its builders disappear tomorrow? Small, deterministic code that everyone can check leaves fewer places to hack and fewer parties to trust. As Ross added in an answer, “Dev needs to focus on simpler code.”
The same instinct appears in Ethereum cryptography, where Vitalik said Ethereum’s roadmap is now “hash-only,” in part because hashes are designed to have no exploitable structure for AI-accelerated mathematics to decipher.
Of course, there is a downside. Public, immutable code is also readable by AI attackers, and you can’t fix it once it’s live. Minimalism therefore works best if the code is small enough to be verified exhaustively.
This is where formal verification, that is, mathematically proving that the code cannot break certain rules, becomes convincing.
Historically, this approach has been extremely costly. Building the seL4 OS core took two person-years, but 20 person-years for its formal verification. Yet in the new research With Theorem released Thursday, Anthropic used Claude to prove that 239 entry points in 123 open source libraries cannot crash on any entry up to 64 MB (under certain assumptions), and it did so in a few weeks.
— (@)
The trap? The proofs evolve with the code. A typical proof of the above effort executed about 600,000 lines for a program of about 18,000 machine instructions, so the work can be tedious. hundreds of millions of AI tokens. Additionally, a proof only covers the properties you specify rather than all possible bugs.
However, notice who benefits the most from cheaper proofs, namely small one-time code, i.e. the exact type of code that minimalists want to put on chain. If AI continues to reduce proof costs, “uncatchable” will matter less, because code increasingly should not be. need patch.
Of course, agility will remain important for large scalable systems like L1s and rollups. Yet in this new AI paradigm, exit testing is more of a security North Star than a simple decentralization ideal. Keep an eye on these trends accordingly.
About the Author
