A report this week from Hacker News flagged a research post from Anthropic in which Claude, the company's large language model, produced a detailed reconstruction of the origins of the elliptic curves used in ECDSA — the Elliptic Curve Digital Signature Algorithm that authenticates the overwhelming majority of modern digital transactions. The Anthropic research post, published at anthropic.com, described how Claude synthesized decades of scattered academic papers, NIST standardization documents, and NSA design memos to trace why specific curves — particularly P-256, also designated secp256r1 — were chosen over alternatives during the late 1990s standardization process.
ECDSA is not a niche concern. The algorithm secures TLS 1.3 connections (the padlock on roughly 95 percent of HTTPS web traffic as of 2026), authenticates software updates on most major operating systems, and underpins nearly every hardware security key sold today. Bitcoin uses a related curve, secp256k1, for wallet signatures. The P-256 curve in particular was adopted by NIST in 1999 and has been mandatory in U.S. federal systems under FIPS 186 ever since.
The tension Claude apparently documented centers on the so-called "seed values" — 160-bit numbers used to generate the curve parameters — whose origins were never fully explained by NIST or the NSA at the time of standardization. Critics including cryptographers Daniel J. Bernstein and Tanja Lange have argued for years that unexplained seed provenance creates the theoretical possibility of a backdoor, because a malicious actor who chose the seed knowing a discrete-log weakness could exploit it silently. Claude's synthesis, per the Anthropic post, did not uncover a smoking gun but did consolidate the documented institutional chain of custody for those seeds more completely than any single prior source had done — pulling from a 1997 draft, a 1999 final standard, and subsequent NIST workshop transcripts.
Anthropic did not claim Claude discovered new mathematical vulnerabilities, and no cryptographer cited in the piece suggested P-256 is currently broken. The research framing positioned the exercise as a demonstration of Claude's ability to reason across long technical document chains rather than as a security disclosure.
For readers who maintain offline or semi-offline systems — backup communications gear, locally signed software repositories, air-gapped document stores — the relevant detail is that ECDSA curve choice is not academic when selecting hardware security tokens or configuring a local certificate authority. The distinction between NIST curves (P-256, P-384) and independently auditable curves like Curve25519 or Ed25519 matters specifically because Curve25519 was designed by Bernstein with fully public, nothing-up-my-sleeve seed derivation: the seed is simply the digits of pi, verifiable by anyone. Many preparedness-oriented self-hosters running local PKI infrastructure have already migrated toward Ed25519 for exactly this reason, and our hardware security key roundup covers which current tokens support it natively. The Anthropic research does not change the underlying math, but it does consolidate the historical record in a form that makes the curve-selection conversation easier to have with less-technical household members or local mutual-aid network administrators who manage shared infrastructure.
