Where did the numbers underlying important digital signatures come from? On October 16, Anthropic published a reconstruction of previously unexplained seeds, achieved by an internal Claude research model. The interesting advance is technical detective work with a checkable result, rather than broken encryption: seemingly suspicious constants acquire a reproducible origin story.
Key takeaways
- The work concerns public elliptic curve parameters, rather than individual users’ secret keys.
- Their origins had long interested cryptographers, and a public bounty was intended to advance the search.
- The finding illustrates how AI can support investigations with clearly verifiable outcomes.
- It gives users no reason to change passwords or encryption settings.
Why public numbers can raise questions
Digital signatures help establish whether data has changed and who it came from. They involve a private key for signing and a public key for verification. The historical NIST document FIPS 186-2, published in 2000, explicitly describes this distinction. ECDSA is its elliptic curve variant. A curve here is a mathematically defined working environment, rather than a drawing whose appearance promises security.
The parameters defining that environment can be public. That makes asking about their origins legitimate: a public procedure can be examined, but an unexplained choice of its underlying numbers leaves an additional question of trust. Knowing a calculation does not automatically explain why its particular input was chosen. A useful analogy is a reproducible recipe: it explains how ingredients produce a dish, rather than where the ingredients came from.
In 2023, cryptographer Steve Weis documented that the NSA had contributed the seeds in question. He collected accounts, historical records, and a recollection that humorous English phrases might have been hashed. A hash maps inputs to a fixed-size value. That story offered a lead for an investigation, rather than an established reconstruction. Weis made gaps in his research visible, instead of treating a missing paper trail as proof of manipulation.
In October 2023, Filippo Valsorda offered a $12,288 bounty, rising to $36,864 if donated to charity. His announcement also emphasized that he did not consider the backdoor suspicions well founded. Even before AI became involved, two questions were therefore distinct: is the system secure, and can its origins be better explained? The second can be interesting without first answering the first in the negative.
An old programming tool leaves a fingerprint
According to Anthropic, the trail led to lrand48 and timestamps from 1997 and 1998. The company says the search took less than half an hour for the five NIST prime curves, including P-256. It does not identify a specific publicly available model version. Questions remain about B-163 and several base points.
The Linux documentation explains lrand48: it returns nonnegative pseudorandom numbers with 31 output bits, using a 48-bit internal state. Initialized identically, such generators produce the same sequence. That is useful for simulations, but it does not make their output suitable for secret cryptographic randomness. Reproducibility alone is not a security guarantee.
The application matters here. A tool whose output can be reconstructed can leave traces in public constants. That does not mean a private signing key has been recovered. Equating the two skips the finding’s actual significance. The interesting direction is the reverse: reproducing an old number sequence can help explain how a publicly known result was created.
For technically minded readers, this offers a useful investigative pattern. Instead of trying ever more possible phrases, ask which software available at the time might have left the observed structure. That moves an investigation from a remembered account toward properties that can be calculated. A plausible story does not automatically become true; it gains a concrete opportunity to fail against existing data.
What this case shows about AI in research
The value of this kind of investigation is particularly assessable when a proposed explanation can be calculated back to known values. That is a more favorable starting point than a persuasive answer without an independent criterion for checking it. For research practice, it suggests using AI where hypotheses can quickly meet hard counterexamples: historical file formats, old software libraries, or reconstructions of documented calculations.
That is our interpretation of the case, rather than proof of general superiority over cryptographers. A reported completion time does not automatically include model development or the specialist community’s earlier work. Nor does a single successful finding say much about how often the same approach fails on other unsolved questions. The recent physics benchmark for AI agents explores the same basic opportunity from another direction: computational assistance becomes useful when results remain open to expert verification.
The specific outlook is encouragingly practical. Old technical standards still contain questions that may yield to a different investigative approach, without requiring new hardware or a new product. AI can supply additional approaches. For readers, the label “mystery solved” matters less than being able to take the explanation and calculation apart. That opportunity is what turns a surprising announcement into a useful research contribution.