- Cryptopolitan
- Posts
- ๐ตโ๐ซ A Not-Even-Public AI Just Cracked a Government-Grade Cipher
๐ตโ๐ซ A Not-Even-Public AI Just Cracked a Government-Grade Cipher
PLUS: Ex-OpenAI researcher hedge fund bets against AI chip stocks

A not even public AI just cracked a government-grade cipher

The company announced on Monday that an unmatured model, Claude Mythos Preview, discovered a functional attack against HAWK, one of the final lattice-based signature schemes in the running for the position. The iteration took around 60 hours of AI research and about $100k worth of compute. It had gone unnoticed in two years of human review.
What We Broke, and What did not
HAWK-256 is a smaller test parameter NIST provides so that researchers can stress-test the design, not the version anyone would deploy. Fortunately for Mythos, HAWK's underlying lattice contained an unused pattern that dropped the target cost of a key-recovery attack from approximately 264 operations to 238, 67 million times lower. For comparison, code released by Anthropic itself recovers a HAWK-256 key in roughly three hours and 42 minutes on a 96-core server.
The versions anyone would realistically deploy, HAWK-512 and HAWK-1024, are unaffected: their attack cost only falls from 2^150 to 2^108, and from 2^288 to 2^182, still very computationally prohibitive. Along with it, Anthropic announced a second finding: that they achieved 200-800x speedup on an attack against seven-round AES, a research benchmark nobody had improved since 2013.
The problem is not the break itself, but it will sound kinda weird, wellโฆ
It's how it was found. Mythos read the literature, performed its own tests, and isolated the error working semi-autonomously within a research scaffold built on Claude Code. Anthropic claims it then verified Mythos completed its job correctly over hundreds of hours, which means the model found the bug more quickly than its own creators could verify.
As Anthropic put it plainly, this field is getting to the point where language models expose too many bugs for normal triage, verification and fix processes to keep pace.
The reason why bitcoin developers are still watching
Bitcoin never used HAWK. Its signatures use (of course) ECDSA and Schnorr over the secp256k1 curve, a completely different construction, so no wallet is impacted. However, Bitcoin's own post-quantum plan, BIP-360, is based on guidelines using ML-DSA and SLH-DSA, the exact two schemes which NIST finalized in 2024 due to them having received greater scrutiny than newer candidates such as HAWK. That decision looks better today than it did a week ago.
The unanswered question is whether the automated use of AI in cryptanalysis eventually produces anything useful with ML-DSA or SLH-DSA as well, and whether it does so before Bitcoin completely migrates from ECDSA.
The high price of waiting too long
According to estimates cited by TheStreet, there are approximately 6.7 million bitcoins worth nearly $600 billion contained in public keys of wallets that could potentially have their private keys revealed by a competent quantum computer first.
A different analysis, Quantum Horizon, gives a 1-in-6 chance of a sufficiently powerful code-breaking quantum computer existing by 2035, and estimates that if such an event were to happen before migration is complete then 2.3 million coins would be unrecoverable.
No one considers Bitcoin's cryptography to be broken as of this writing. What perhaps changed this week is confidence in how much runway remains to fix it before it breaks.
POLL: Do you think Quantum computing will be able to break Bitcoin's encryption before 2035 |

1๏ธโฃ Binance Goes After Kalshi & Polymarket.
The exchange will apply for a CFTC prediction market license in August, CEO Stephen Gregory said at the Rare Evo conference, the latest move by the firm as it attempts a comeback built on lower fees and new products following years of declining US volume. Gemini already has its license.
Coinbase already partners with Kalshi. Binance.US is several years too late to a market that saw $27 billion in World Cup betting alone just for Kalshi.
2๏ธโฃ Memory chips solely carried Samsung's whole quarter.
KRW 89.5 trillion operating profit, almost all of it from HBM and server DRAM driven by AI demand, mobile was a loss.
SK Hynix and Micron released similar blowout numbers a few days apart. This is, of course, not good news for anyone looking to buy a phone anytime soon, but on the other hand great news for these three. Indeed TrendForce predicts DRAM prices will increase another 58-63% next quarter.
3๏ธโฃ Microsoft surges 8% as Meta tumbles 10% after sharply contrasting AI earnings
Both Microsoft and Meta both reported the same day, and split hard. Microsoft surpassed every expectation and rose 8%. Meta beat on revenue but missed earnings badly on legal and layoff charges, down about 9%. Similar level of AI spending, same quarter, different call on who is actually converting it into profit.
Are you watching?
Ex-OpenAI researcher hedge fund bets against AI chip stocks

Situational Awareness, a hedge fund specializing in artificial intelligence that was established by Leopold Aschenbrenner, a previous employee of OpenAI, is seeking to get additional funding from investors as semiconductor and AI stocks plummeted. The timing is important because this hedge fund has the biggest disclosed bet against the companies causing the drop in the stock prices of semiconductor companies.
POLL: We cover AI every week inside Cryptopolitan Daily. Would you want a dedicated daily AI newsletter from us? |
In case you missed it
๐ Our most read story from last edition: Jensen Huang breaks down AI job fears as big companies return to hiring
Tweet picks by our Dog

Meme of the day
Just for you
Join the Conversation!
We'd love to hear your thoughts and comments. Join our community and stay updated with the latest trends and discussions in crypto.
Twitter | Instagram | Telegram Channel | Linkedin | Facebook
