On March 15, 2025, a large language model named Claude finished a job that most cryptographers would call a career highlight. It found a functional weakness in a post-quantum digital signature scheme. The run took 60 hours. The cost? Under $10,000 in API fees. The result? A valid signature that shouldn’t have existed.
Code doesn't lie. But implementations do.
The event hit my radar the same way the Terra collapse did – as a signal that the narrative is always behind the mechanism. Everyone is worried about quantum computers breaking Bitcoin’s ECDSA. But the real threat might be AI finding holes in the replacements before they even ship. This is not a theoretical exercise. A human named Amir took the output from Claude and generated an obvious invalid signature. That means the weakness is exploitable.
Let’s cut through the hype and look at the stack.
Context – What Actually Broke
The post-quantum digital signature schemes in question are the NIST-standardized candidates: CRYSTALS-Dilithium and FALCON. These are based on lattice cryptography. Their security relies on the hardness of short vector problems, not on factoring or discrete logs. They are designed to be resistant to quantum attacks. But that resistance is only as strong as the code that implements them.
Claude didn’t break the math. It found a flaw in a specific implementation. The analysis team behind the discovery confirmed that the weakness was in the engineering layer – likely a failure to use constant-time operations, a mishandling of randomness, or a memory corruption that leaks the private key through side channels. This is exactly the type of bug that automated scanners miss because they don’t understand cryptographic flow.
I know that pattern from personal experience. In 2020, while auditing the Uniswap V2 factory contract, I found an integer overflow in the liquidity token minting logic. The automated scanners passed it. They checked for arithmetic errors but didn’t audit the logical sequence. I spotted it because I read the raw bytecode. Claude did something similar – it read the human-readable source and traced the execution path until it found a divergence from the specification.
The difference is time. A full-time security team might take two weeks to find a bug like that. Claude did it in 60 hours. The API cost for that kind of continuous inference – roughly 1.4 million output tokens – is between $2,000 and $10,000. That’s one-twentieth the cost of a human auditor. The implications for product security are obvious. The implications for blockchain security are tectonic.
Core – Order Flow Analysis of the Real Threat
Let’s map this onto the crypto market. Every Layer 1 and major DeFi protocol is planning a post-quantum migration. Ethereum’s roadmap includes a transition to quantum-resistant signatures. Bitcoin has a BIP for Schnorr, but full quantum resistance is a longer game. The migration is slow because the math is complex and the implementations are unproven. Now we know that even the best-funded standards have bugs that an off-the-shelf LLM can find for $10,000.
The attack surface is massive. Imagine a protocol deploys a new signature scheme for staking or governance. A single implementation bug could allow an attacker to forge signatures on behalf of any user. That’s not a theoretical risk – it’s a direct threat to delegated voting, multisig wallets, and cross-chain bridges. The attacker doesn’t need to break the lattice problem. They just need to find the same bug Claude found.
And here is where the narrative breaks from reality. The market will interpret this discovery as "AI is now a cryptanalyst" and pour money into anything with "AI security" in the name. But the actual insight is far more boring and far more important: the bottleneck in cryptographic security is not the math, it’s the code. And AI makes code auditing cheaper, which lowers the barrier for both defenders and attackers.
Earlier this year, I audited an AI-driven trading bot that claimed 30% monthly returns. I found it was just executing high-frequency, low-margin trades on DEXes, burning gas on every swap. The returns were an illusion created by a cherry-picked backtest. I shorted the token after publishing the data. That case taught me a rule I apply here: if you can’t verify the mechanism, don’t buy the narrative.
I audit the logic, not the hope.
The Claude discovery has no direct revenue attached. It doesn’t change the economics of any token. It doesn’t create a new moat for Anthropic – competitors like OpenAI and Google DeepMind can replicate the approach with similar models. The only durable advantage is the dataset of known implementation weaknesses, which Anthropic can build over time. That’s a lead measured in weeks, not years.
Contrarian – The Market Will Misread This
The conventional take is that this is a win for security. AI helps find bugs faster, so we can patch them before deployment. That’s true, but it’s only half the picture. The other half is that the same tool can be used by criminals to find zero-day exploits in deployed systems. The speed of discovery compresses the window for fixing. In a bull market, projects rush to launch with minimal auditing. A malicious actor could run Claude against a live protocol, find a bug, and drain the pool before the team even finishes their security review.
Speed is the only shield in a flash loan. And now it’s the only shield in signature forgery.
Second, the hype will create a false sense of security. Projects will claim "we passed the Claude audit" as a marketing badge. But the test was run on one specific implementation with a known prompt structure. The reproducibility is low. Run the same task with a different RNG seed or a different model version, and you may get a different bug – or none at all. The confidence rating from the original analysis was C (medium). That’s a coin flip. Betting a protocol’s security on a single LLM run is as reckless as betting your portfolio on a single oracle.
Third, this event will slow down standardization. If NIST sees that even the top candidates have implementation weaknesses, they may delay final approval. That freezes the migration timeline. Every month of delay is another month that classical cryptography remains vulnerable. The irony: the AI discovery that should accelerate fixes might instead stall them.
Takeaway – Verify the Exit
I’m not saying sell everything and run to cash. I’m saying the market is pricing in a fantasy – that AI will make crypto safer. What it actually does is make the game faster. The winners will be the protocols that treat their signature implementations as live security boundaries, not as afterthoughts.
Trust the stack, verify the exit.
If you hold tokens in a project that claims quantum resistance, demand three things: the exact library they use for signatures, the commit hash of the audited code, and a report from an independent AI-driven audit. If they can’t provide it, you are betting on hope. And hope doesn’t pay out in bear markets.
The real race is not between humans and AI. It’s between AI-assisted auditors and AI-assisted attackers. Both sides now have the same weapon. The difference is who fires first.
I’m watching the order books for the first project that gets drained because they skipped the AI check. That will be the signal to rotate out of every post-quantum narrative token and into the protocols that actually run audits. The code doesn’t lie. But it will if you don’t verify.
