The headline said the mystery was unpicked. The third paragraph said the mystery remains. Both describe the same event, and only one of them is defensible.
Claude โ Anthropic's model โ reportedly reconstructed the preimage string that hashes into the seed used to generate NIST P-256, the elliptic curve otherwise known as secp256r1. The stated method: a "timestamp recipe." The stated result: additional provenance evidence. The stated limit, easy to miss in a scrolling feed: the selection criteria and the origin of the base point remain unexplained.
That is the entire forensic payload. No paper. No repository. No independent reproduction. No named author. What circulated was a vendor claim, filtered through a media summary, filtered again through an aggregator blurb carrying a newsletter promotion. Three levels of transcription. Zero levels of verification.
Reconstructing the timeline of a seed's birth is a legitimate investigative exercise. Calling it a resolution is not.
Context: two curves, one convenient confusion
Ethereum accounts sign with secp256k1. That is a Koblitz curve. It has nothing to do with P-256.
Passkeys โ WebAuthn, Touch ID, Face ID, the secure enclave in your phone โ sign with P-256. Different curve, different trust model, different provenance.
The bridge is EIP-7212, and its rollup analogue RIP-7212: a precompile that makes secp256r1 verification cheap enough to execute on-chain. Without it, verifying a Passkey signature costs enough gas to make smart-account UX uneconomical. With it, a biometric signature can authorize an ERC-4337 account without a mnemonic ever existing.
So the "Ethereum Passkey security" framing is not wrong. It is compressed past the point of accuracy. Anyone reading "P-256" as "Ethereum's native curve" has already lost the thread, and that misreading is doing quiet work in most of the commentary I have seen this week.
Why does anyone suspect P-256 at all? Because NIST published the P-curves with seeds described as "verifiably random" โ the SHA-1 hash of a string that was never disclosed. The seed itself is public; it is printed in the standard. The preimage was withheld. Combine an undisclosed input with the Dual_EC_DRBG backdoor episode, and you get two decades of entirely reasonable suspicion. Not proof. Suspicion.
Core: what was actually recovered
Here is the mechanics, stripped of narrative.
The seed in the standard is a hash output. If the preimage is a short string in an enumerable format โ a date, a timestamp, a fixed-prefix identifier โ then the search space is finite and traversable. Recovering it is a search problem, not a cryptanalytic one.
That distinction matters enormously. Breaking a curve means finding collisions, inverting the discrete logarithm, or exploiting a structural weakness in the group law. Recovering a preimage from a small candidate set means writing a loop and letting it run. The genuinely interesting claim is methodological: a language model was used as a heuristic to propose candidate formats, and deterministic enumeration was left to do the actual work. If that is what happened, it is a real demonstration of AI as a research instrument.
Decoding the algorithmic chaos of a provenance search is not the same as decoding a cipher, and the difference is the whole story.
Based on my audit experience, the most expensive errors in this industry are not code bugs. They are category errors โ mistaking a documentation artifact for a security proof. I watched this pattern in 2017, when I built ETL pipelines across more than 500 ICO token distributions and found that 70% of "community-driven" pre-sales were controlled by fewer than ten wallets. The contracts were fine. The disclosures were technically fine. The narrative was the vulnerability. The same structural error is available here: a recovered preimage is a disclosure improvement, not a soundness proof.
Decoding the forensic silence of a vendor claim means asking what was not said. The source material does not say the curve is safe. It does not say the selection criteria were benign. It says provenance evidence was supplemented. That restraint is the most credible sentence in the entire story โ and it is precisely the sentence that gets deleted in retransmission. I have run enough data pipelines to know that the missing field is the finding. When a distribution table has an empty column labeled "team allocation," that emptiness is the analysis.
Contrarian: the trust gap does not close
Suppose the preimage is confirmed and reproduced tomorrow. The gap remains open, for a structural reason that no amount of compute closes.
Curve parameter generation is not a single act. It is a selection over a space of candidates. A generator can compute thousands of seeds, inspect the resulting curves, and publish the one that suits them. Recovering the preimage for the published seed proves only that this preimage hashes to this seed. It says nothing about how many candidates were discarded, or why.
Verifiable randomness narrows the space of possible manipulation. It does not eliminate it. This is exactly where the coverage will fail โ the story will be read as exoneration, when the technical content is a narrowing of suspicion.
And here is what the narrative buries completely: the curve was never the live risk. The live risks sit downstream โ authenticator firmware, precompile implementation correctness, and the signature verification logic inside account-abstraction wallets. A philosophically uncomfortable provenance has never drained a user's balance. A mis-implemented verification path does it routinely.
Takeaway: what to watch
Three signals over the next two weeks. First, whether Anthropic publishes a reproducible script or technical note โ disclosure upgrades the claim from assertion to evidence; silence should be discounted to zero. Second, whether EIP-7212 precompile implementations attract independent audit attention as a direct consequence. Third, whether any major account-abstraction wallet publishes its secp256r1 verification path.
In a chop market, where positioning is the only edge available, this is precisely the kind of story that consumes attention and returns nothing tradeable. The question worth asking is not whether the mystery was solved. It is why a search over a timestamp format is being reported as a verdict on the curve that secures a billion biometric logins.