Claude cracks NIST ECDSA seeds, Anthropic junior hires fall 27% to 9%

TL;DR

• Claude traces NIST’s ECDSA seeds to lrand48 and a 1997-98 Unix timestamp.
• Companion paper claims 17× Toffoli-gate cuts for modular inversion, on paper only.
• Valsorda reproduces the recipe in Go and triples his bounty to $36,864.
• Junior hires fell 27% to 9% of Anthropic’s intake since early 2025.
• Google ATLAS finds full automation in under 10% of 15M interactions, capping extrapolation.

Two unrelated Anthropic stories carry today’s research beat, and the stronger artifact is the crypto one. Internal Claude traced NIST’s P-256 ECDSA curve seeds back to lrand48 plus a 1997–98 Unix timestamp — a 25-year-old mystery — and Filippo Valsorda reimplemented the recipe in Go for byte-for-byte confirmation, tripling his bounty to $36,864. The companion quantum-circuit paper claiming a 17× Toffoli-gate reduction is softer: on-paper only, 12% more qubits, no hardware run.

The second feature turns Anthropic’s own headcount into the dataset. Junior technical hires fell from 27% to 9% of intake since early 2025, and interactive agent hours now outnumber human hours 6:1 internally. The external checks cut the other way here — Google’s ATLAS study of 15M interactions finds full automation in under 10% of cases, and Acemoglu caps AI’s decade work-replacement share at 5%. Internal numbers are real; the extrapolation outward is where today’s tension sits.

Claude solves NIST curve mystery, shrinks quantum circuits 17×

Source: anthropic-research · published 2026-10-16

TL;DR

• Anthropic’s internal Claude traced NIST’s ECDSA curve seeds to lrand48 plus a 1997–98 Unix timestamp.
• A companion paper claims order-of-magnitude quantum-circuit gate reductions from the same model configuration.
• Filippo Valsorda reimplemented the ECDSA recipe in Go, confirmed byte-for-byte seed reproduction, and tripled his bounty to $36,864 for ISRG.
• The 17× Toffoli-gate reduction for modular inversion costs 12% more qubits and is an on-paper estimate — no hardware run.
• NIST skipped 966 viable P-256 candidates before accepting #1,842, leaving room for selection-bias backdoor theories.

The twin drop

Anthropic put out two papers on the same day to make one argument: Claude, in an unreleased internal configuration, can produce original scientific results. One tackled a 25-year cryptographic mystery; the other claimed order-of-magnitude improvements in quantum circuit design. Read together, they are a pitch — the “autonomous scientist” framing — that lands unevenly.

ECDSA: verified, with loose ends

The strong result is the cryptographic one. Claude recognized the fingerprint of the lrand48 linear congruential generator (top 31 bits of a 48-bit state, predictable zero bits), applied LLL lattice reduction to recover truncated internal states from NIST’s 20-byte seeds, and reproduced the seeds for P-192 through P-521 by linking them to specific Unix timestamps from 1997–1998. The initialization state’s low 16 bits were always 0x330E — the libc default. The seeds were hidden in plain sight.

The strongest validation didn’t come from Anthropic. Filippo Valsorda, who posted the original $12,288 bounty in 2023, reimplemented the recipe independently in Go, confirmed exact reproduction, and donated the tripled $36,864 pot to the Internet Security Research Group ¹. That is as close to community sign-off as cryptography gets.

The backdoor debate, however, is not quite dead. For P-256 the first viable seed appeared at candidate 876, but NIST’s designers skipped 966 more before accepting #1,842 — a gap with no documented justification ². P-224 passed over three viable options. Combined with roughly 30 bits of timestamp freedom, an adversary who already knew a rare weak-curve class could in principle have steered selection while preserving plausible deniability. The B-163 binary-curve seed doesn’t fit the pattern at all, the P-curve base points remain untraced, and Claude flagged an “extra rejection criterion” beyond the published ANSI rules ³. Call it 90% resolved.

Quantum circuits: louder claim, thinner validation

The companion paper is where the “autonomous scientist” framing strains. John Preskill called Claude’s resource reductions genuinely surprising, conceding some projects “would not work” without AI-assisted optimization — then immediately noted that gate-count compression doesn’t touch the dominant bottleneck of logical-qubit error rates ⁴. The headline 17× Toffoli reduction for modular inversion traded away 12% more qubits and remains a paper estimate; nothing has run on hardware ⁵. There is no Valsorda-equivalent independent verifier.

The capability-gating subtext

A smaller thread in the response coverage is worth naming. The internal Claude was permitted advanced offensive cryptanalysis; public API users routinely hit refusals on comparable tasks ⁶. Developers noticed. If Anthropic wants “autonomous scientist” to mean something outside its own lab, that asymmetry is the next thing to resolve — and it’s a policy question, not a model one.

Further reading

• Designing more efficient quantum circuits with Claude — anthropic-research

---

Anthropic’s junior hiring share collapsed from 27% to 9%

Source: anthropic-research · published 2026-10-16

TL;DR

• Junior technical hires fell from 27% to 9% of Anthropic’s intake since early 2025, with headcount share dropping 21%→14%.
• Interactive agent hours outnumber human hours 6:1 inside Anthropic as of August 2026, with code commits per engineer up 5×.
• Google’s ATLAS study of 15M interactions shows full automation in under 10% of cases, incompatible with Anthropic-scale adoption being typical.
• Acemoglu estimates AI replaces ≤5% of work this decade, calling Anthropic-style extrapolation implausible.

The drop, not the data point

Anthropic published a coordinated three-part workforce bundle on October 16 — internal employment telemetry, a public-vs-practitioner expectations survey, and an umbrella synthesis — explicitly pitching itself as a “leading indicator” for the broader labor market. The framing is the news. Any single statistic (6:1 agent hours, 5× commits, 22% meta-work) is a data point from one Claude-maximalist firm. The three pieces together are an argument that the Anthropic workforce is the near-future workforce, and that policymakers should price that in now.

That argument deserves to be split in two, because the evidence behind it does.

What survives independent scrutiny

The junior-hiring collapse travels. Anthropic’s internal share of new technical hires in junior roles fell from 27% to 9%; Stanford’s 2026 AI Index, drawing on ADP payroll data, finds employment for developers aged 22–25 dropped ~20% between 2024 and 2026 while developers 30+ kept growing ⁷. The shape matches: firms aren’t firing, they’re “failing to hire.” When two very different data sources — one firm’s HR system and nationwide payroll — converge on the same demographic cliff, the signal is real. Anthropic’s leading-indicator claim is strongest here, and it’s also where the policy problem is hardest: the company itself concedes private action alone won’t fix it.

What doesn’t travel outside Anthropic’s walls

The productivity story is a different matter. Google’s ATLAS study — 15 million de-identified Gemini interactions — reports AI appears in 68% of occupations but is applied to only ~20% of tasks, with fewer than 10% of interactions producing full automation ⁸. That is incompatible with a 6:1 agent-to-human-hours regime being anywhere close to typical. Daron Acemoglu’s public response put a number on it: AI will replace at most 5% of human work this decade and add ~1.5% to GDP, bottlenecked by organizational inertia and the “last 1%” of oversight ⁹.

The “5× commits, no change in code slop” claim takes the sharpest practitioner fire. Reviewers point out that agent workflows fragment one coherent human change into many small commits, so commit-count inflation is partly workflow artifact ¹⁰. Worse, AI-generated tests “embed existing bugs as testable patterns” — testslop that Anthropic’s own slop-rate metric is structurally blind to ¹⁰. Independent analysts have been blunt that the report is a “single-company self-portrait” grading its own homework with the metrics most flattering to Claude ¹¹.

The policy fight the bundle is really about

Widespread stock ownership is crucial for pro-worker AI adoption.

That line from Anthropic’s synthesis is the real pitch — displacement as a compensation-design problem, solved by equity. The AFL-CIO’s Workers First Initiative treats the same facts as a bargaining problem and rejects corporate self-regulation outright, with Liz Shuler arguing unions are the only institutions that can enforce guardrails against algorithmic management ¹². Both sides now agree the junior-cohort data is real. They disagree about who captures the gains, and Anthropic’s bundle is a well-timed opening bid in that fight — not a neutral measurement.

Further reading

• The AI expectations gap — anthropic-research
• Working at the frontier shapes expectations for the future — anthropic-research

Footnotes

1. Filippo Valsorda — words.filippo.io ‘seeds-bounty’ — https://words.filippo.io/seeds-bounty/

   Valsorda independently reimplemented the identified algorithm in Go, confirmed the reconstructed process perfectly matched the historical seeds, and awarded the tripled bounty of $36,864 to the Internet Security Research Group (ISRG).

   ↩

2. deniz.in commentary (HN-style write-up) — https://deniz.in/anthropic-says-claude-traced-nist-ecdsa-curve-seeds-to-lrand48-and-unix-timestam

   For the P-256 curve, the first valid candidate appeared at attempt 876, yet the designers skipped nearly a thousand more before selecting candidate 1,842 — this unexplained gap suggests candidates were being discarded, potentially to find a seed with ‘nothing-up-my-sleeve’ optics that still possessed a secret weakness.

   ↩

3. madrobot.blog technical recap — https://madrobot.blog/2026/10/16/claude-nist-ecdsa-curve-seeds-nsa-mystery-solved-lrand48/

   The seed for the binary curve B-163 did not match the pattern, and the specific ‘base points’ for the P-curves still lack a clear derivation… Claude also observed an unexplained ‘extra rejection criterion’ used alongside standard ANSI rules.

   ↩

4. MindStudio — Preskill on OpenEvolve — https://www.mindstudio.ai/blog/john-preskill-oatomic-openevolve-llm-quantum-algorithm-optimization

   John Preskill remarked that the magnitude of resource reduction was unexpected, noting that the project simply ‘would not work’ without AI-assisted optimization — while cautioning that achieving the error rates for logical qubits remains a monumental engineering hurdle gate reduction alone cannot solve.

   ↩

5. Unite.ai — quantum circuits coverage — https://www.unite.ai/anthropic-says-claude-beat-best-published-quantum-circuit-designs/

   The 17x Toffoli-gate reduction for modular inversion required 12% more qubits as a tradeoff, and the result remains an on-paper resource estimate rather than a demonstration on physical quantum hardware.

   ↩

6. KuCoin news flash — https://www.kucoin.com/news/flash/anthropic-s-claude-solves-25-year-elliptic-curve-seed-mystery-bitcoin-avoids-nist-curve

   Bitcoin avoids NIST curve — the discrepancy between Anthropic’s internal models being permitted to perform advanced cryptanalysis while public users face ‘capability gating’ or refusals on similar tasks drew concern from developers.

   ↩

7. Forbes — Josipa Majic — https://www.forbes.com/sites/josipamajic/2026/08/09/coding-jobs-vanish-for-juniors-as-ai-reshapes-career-path/

   Stanford’s 2026 AI Index Report, utilizing ADP payroll data, found that employment for developers aged 22–25 plummeted nearly 20% between 2024 and 2026… headcount for developers aged 30 and older continued to grow, suggesting companies are not firing existing staff but are simply ‘failing to hire’ the next generation.

   ↩

8. Unite.AI on Google’s ATLAS study — https://www.unite.ai/googles-own-data-shows-ai-adoption-is-broad-but-shallow/

   While AI use appears in 68% of occupations, workers typically apply it to only 20% of their tasks… fewer than 10% of interactions result in full task automation, framing agents as collaborative assistants rather than autonomous replacements.

   ↩

9. FourWeekMBA (summarizing Acemoglu response) — https://fourweekmba.com/ai-91-of-anthropic-staff-expect-a-nobel-level-ai-by-2030/

   Nobel laureate Daron Acemoglu… argued that AI would likely replace at most 5% of human work over the next decade, adding only about 1.5% to total GDP… organizational inertia and the ‘last 1% problem’—where human oversight remains essential for accuracy—will slow adoption far more than the technology’s theoretical capabilities suggest.

   ↩

10. Hacker News discussion — https://news.ycombinator.com/item?id=50059523

    ‘commit volume’ is a poor proxy for value. Critics argue that agents often generate dozens of minor, fragmented commits that would have been a single cohesive change for a human… AI-generated tests often ‘embed existing bugs as testable patterns,’ leading to a state of ‘testslop’.

    ↩ ↩²

11. BiggoFinance analysis — https://finance.biggo.com/news/f606a952-e433-473e-bca6-104a8b0bc27b

    Critics have labeled the report a ‘single-company self-portrait,’ arguing that Anthropic is essentially ‘grading its own homework’ by showcasing its own product’s (Claude) impact on its own workflows… ‘code commits’ as a productivity metric [is] ‘notoriously imperfect’.

    ↩

12. AFL-CIO (Liz Shuler) — https://aflcio.org/issues/future-work/ai

    AFL-CIO President Liz Shuler has argued that unions are the only trusted institutions capable of enforcing ‘guardrails’ against workplace surveillance and algorithmic management… remains critical of corporate ‘self-regulation’.

    ↩