CoreWeave, Anthropic, & Debt Fears Walk into a Bar

TLDR;

Another week, another stressful effort to pack everything that happened into one newsletter. But I think I did pretty well this time. From AI’s latest discovery to CoreWeave’s flirting with self-implosion, this week had it all, but I’m afraid mostly on the negative side.

Sometimes we win, sometimes we lose, as they say. But this past week AI is accumulating many losses; perhaps one too many.

Enjoy if you can.

CRYPTOGRAPHY

Decades-old mystery solved by AI in half an hour

Anthropic said on October 16 that an internal Claude research model had uncovered the origins of numbers used in widely adopted digital-signature standards, resolving much of a cryptographic mystery that had persisted for more than 25 years. Incredibly, the company says the model found the answer in under half an hour.

To understand this, we need to do a quick review of how modern data is protected. Modern communications today are secured by public-private keys. Each of us has a public key, known to the rest of the network, and a private key connected to that public key that only we know. The key is that both keys can encrypt and decrypt each other’s ciphertexts. Or, to put it differently, only the paired key can decrypt the other’s ciphertext.

And what’s the point? Well, because it lets us do two things: digital signing and encryption, the basis of all modern security.

Digital signatures help computers verify who signed a message and whether it was modified. This matters because you want to verify you’re receiving data from who you think you are. To do this, first, a mathematical function hashes the message into a fixed-size fingerprint. The sender then uses their private key to sign that fingerprint, allowing others to check the signature using the corresponding public key.

Encryption ensures only the recipient can see the original contents. The sender takes the recipient’s public key, uses it to encrypt (i.e., scramble) the data, and sends it. This way, only the recipient can decrypt it using its private key.

These two procedures secure most Internet data because they protect communication from tampering or theft.

And for the first, a popular method is ECDSA, a widely used signing system that performs its calculations using mathematical curves. Several standard curves were generated from starting numbers supplied by the NSA without an explanation of where those numbers came from. That secrecy prompted concerns that the agency might have deliberately selected curves with hidden weaknesses.

But now, Claude has identified patterns matching an ordinary 1990s pseudorandom-number generator, then reconstructed its initial inputs: timestamps from the period when the curves were created. Running the generator again with those inputs reproduces the published starting numbers exactly.

What an absolute banger from Anthropic.

TheWhiteBox’s takeaway:

The big takeaway is that Anthropic says this strengthens confidence that the curves were not deliberately weakened, although some details remain unexplained, including why certain apparently suitable candidates were skipped.

Still, another incredible exposé of Anthropic’s models' capabilities. That said, remember that this…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.