Anthropic says its Claude Mythos Preview model found a serious weakness in HAWK, a candidate post-quantum signature scheme that had already survived heavy human review. That is good news for security research, and a blunt reminder that quantum-safe cryptography is still a work in progress, not some magic shield.
- Claude found a flaw in HAWK, a proposed post-quantum digital signature scheme.
- The attack cut the estimated work factor for HAWK’s smallest parameter set from 2^64 operations to 2^38.
- Anthropic says Bitcoin and Ethereum are not affected; the issue is specific to HAWK.
- The bigger signal: AI-assisted cryptanalysis is becoming a real force, and that cuts both ways.
Anthropic says the HAWK attack took roughly 60 hours and around $100, 000 in API cost. That is not pocket change, but it is also not science fiction. Frontier AI is increasingly able to help cryptographers find weaknesses before those weaknesses get baked into systems people actually depend on.
HAWK is a proposed post-quantum digital signature scheme. In plain English, digital signatures prove that a message or transaction was authorized and not tampered with. If signatures fail, the whole trust model starts to wobble. That is not a small bug. That is the plumbing under the house cracking.
Anthropic says the attack reduced the effective key strength of HAWK’s smallest parameter set from an estimated 2^64 operations to 2^38. That does not mean HAWK is “broken” in the dramatic, internet-brained sense. It means the security margin got cut down hard enough to matter. In cryptography, security margin is the difference between “comfortable” and “call the fire department.”
The important nuance: this does not mean Bitcoin or Ethereum are suddenly in danger from HAWK. Anthropic says the weakness is specific to HAWK and does not affect deployed Bitcoin or Ethereum cryptography. So no, your coins are not being vaporized by a flaw in some unrelated post-quantum candidate.
That distinction matters because “quantum risk” gets thrown around like a catch-all curse word. It is not one giant blob of doom. A weakness in one candidate scheme does not mean all cryptography is toast. It does mean the review process for future standards needs to be ruthless, because apparently the math is not interested in your roadmap deck.
Anthropic’s broader point is harder to ignore: AI can shorten the time available to migrate to stronger security. The company also reported separate gains against a deliberately weakened version of AES, the Advanced Encryption Standard, where Claude initially thought the task was too hard before succeeding after additional prompts. Anthropic says that work improved the attack by 200 to 800 times, depending on the setup.
Against Poseidon, a hash function widely used in zero-knowledge proof systems, blockchain rollups, and privacy-focused protocols, the model achieved less than a tenfold improvement. That uneven result is the real signal. AI is not magically cracking everything. It is useful in some places, less useful in others, and still heavily dependent on prompting, scaffolding, and brute-force compute.
Anthropic said the AES work generated nearly one billion output tokens over three days, with human researchers then spending almost a month verifying the results. That is a very expensive way to learn an old lesson: the machine is not replacing the cryptographer. It is becoming a loud, fast assistant that can surface things humans miss, and sometimes surface them before the bad guys do.
For Bitcoiners, the HAWK finding matters less than what it says about the coming migration away from legacy signatures. A draft Bitcoin Improvement Proposal, Post Quantum Migration and Legacy Signature Sunset, argues for a post-quantum migration path and a phased sunset of legacy ECDSA/Schnorr signatures. In other words, the network’s long-term security conversation is not about whether change is needed, but how painful and how rushed that change will be.
That proposal is still draft material, not final consensus law. But the direction is obvious. If quantum computing keeps advancing and cryptanalytic tooling keeps getting sharper, leaving old signature schemes in place forever is not a strategy. It is a liability with a nice logo.
The good side of this development is straightforward: AI can help expose weak designs earlier, before they are trusted with real money and real infrastructure. That is exactly what public review is supposed to do. Find the hole now, not after the hole is monetized.
The darker side is just as obvious. If AI can help defenders stress-test schemes faster, it can also help attackers do the same. That shrinks the window between “looks safe” and “actually safe.” Anyone selling quantum-safe security as if it were a solved problem is peddling fantasy. The standards are still being hardened in public, and the public part is doing exactly what it should: embarrassing bad assumptions.
Anthropic frames the HAWK result as responsible cryptographic stress-testing, not a practical break of deployed software. That framing is fair. HAWK is a candidate scheme under review, and candidate schemes are supposed to get punched in the mouth before they get adopted.
Still, the takeaway is bigger than one algorithm. AI is moving into the guts of security analysis, not just code generation and chatbot theater. That should make defenders more capable, attackers more dangerous, and every “we’re already quantum safe” claim sound a lot more suspicious.
Key takeaways
-
Does this break Bitcoin or Ethereum?
No. Anthropic says the HAWK weakness is specific to HAWK and does not affect deployed Bitcoin or Ethereum cryptography. -
How serious is the HAWK flaw?
Serious enough to matter for cryptographic review and standardization, but not a practical break of deployed systems. It weakens the scheme’s security margin rather than wiping it out. -
Why does AI finding cryptographic bugs matter?
Because it can speed up both defense and offense. That can help remove weak designs earlier, but it also shortens the time before attackers may catch up. -
What does the 2^38 figure mean?
It is an estimate of attack complexity. Lower numbers mean weaker security, so the reduction from 2^64 to 2^38 is a meaningful downgrade. -
Should Bitcoin users worry about quantum computers right now?
Not because of HAWK. But long term, yes, which is why post-quantum migration proposals like BIP-361 matter and deserve serious attention.
Anthropic AI Finds Cryptographic Flaw, Raising Questions about how much trust should be placed in “reviewed” crypto before AI starts stress-testing it at scale. The company’s cryptanalysis work is useful, but it also exposes how fragile the assumption of human-only review really is.
Anthropic’s result is a useful reality check. Post-quantum cryptography is not a finished product sitting safely on a shelf. It is an ongoing contest, and AI has now entered that contest in a meaningful way. That can make security stronger. It can also make complacency a lot more expensive.
For those tracking the broader standards race, Post-Quantum Cryptography: Additional Digital Signature work remains part of the long slog toward replacing legacy schemes with something less brittle under future attack models. And yes, this is exactly the kind of boring-looking standards grind that decides whether your future money infrastructure is sturdy or a house of cards with good branding.
That standards process sits inside the larger NIST Post-Quantum Cryptography Standardization effort, which is basically where the sausage gets made for post-quantum security. The annoying part is that sausage-making takes time. The useful part is that time is cheaper than a total cryptographic faceplant.
There is also a live debate around whether some privacy systems are actually as future-proof as advertised. The Zcash community has asked, Is Zcash Actually Quantum Private? a question that matters because “private today” and “private against future quantum threats” are not the same thing. Plenty of projects love the first slogan. Fewer can honestly defend the second without sweating a little.
That same skepticism is why analysts continue to revisit assumptions around major Bitcoin upgrade paths. As Samson Mow Warns Bitcoin’s Post-Quantum Upgrade Could Break, the migration problem is not just about quantum threat models, it is about the real possibility of implementation mistakes, rushed governance, and self-inflicted wounds. The hardest bugs are often the ones humans schedule for themselves.
For readers wanting a sharper technical critique of the cryptography itself, Some thoughts about Anthropic's new cryptanalysis results offers a useful counterbalance to the hype. That is the proper move here: praise the discovery, then check the math twice, because the crypto world has a long and embarrassing history of confusing “promising” with “done.”
Further reading
A related take on AI and crypto market turbulence worth a look:
