The Intake
The Intake — Saturday, October 17, 2026
On the substrate
Agent traces NIST elliptic curve seeds to 1990s RNG, closing a 25-year open question
Anthropic Frontier Red Team Filippo Valsorda (bounty holder) Unite.AI
If your stack relies on HTTPS, JWT signatures, or elliptic-curve cryptography, you've been running on NIST curves whose provenance was an open question until this week. An Anthropic Frontier Red Team post, published October 16, 2026, reports that a Claude research model traced the seeds for five NIST elliptic curves to the C standard library's lrand48 random-number generator. The five curves: P-192, P-224, P-256, P-384, and P-521. The seeds were constructed from Unix timestamps ranging from May 1997 through April 1998.
The model identified a characteristic zero-bit pattern in lrand48 output. The generator returns only the top 31 bits of a 48-bit internal state. That truncation produces the zero-bit pattern at a probability of 1 in 5,810. The model recovered the generator's internal state via LLL lattice reduction. It then reconstructed the full seed-generation sequence in under 30 minutes. The team included John Wood, Henry de Valence, Ava Howell, Zack Voase, and Nicholas Carlini at Anthropic, among others.
Filippo Valsorda, who held an open $12,288 bounty for the solution, confirmed the result. Additional contributions had tripled the prize to $36,864. The full amount will be donated to the Internet Security Research Group, which operates Let's Encrypt. Valsorda's stated conclusion: the generation process's simplicity makes deliberate backdoor introduction implausible. If you've been treating "no known backdoor" for P-256 as an assumption you couldn't verify, this result gives you the verification: a standard C library and a set of timestamps, nothing more.
---
Claude designs quantum circuits beating published benchmarks on 7 of 8 tasks
Anthropic Research arXiv 2610.22367 The Quantum Insider
If you've been running closed-loop agent workflows — propose, score, revise — Anthropic's October 15 preprint gives you a scale reference for what that pattern produces on a constrained optimization problem with verifiable ground truth. The setup: Claude Mythos 5 proposed Python implementations of quantum circuits. A circuit evaluator scored each on Toffoli gate count and qubit requirements. The model revised based on that feedback.
The preprint reports the model surpassed previously published circuit designs on 7 of 8 tasks. The strongest single result was a quantum decoding circuit. Its Toffoli gate count fell from 5.1 million to 440,000 — approximately 12×. The peak qubit count for that circuit dropped from 1,938 to 1,837. An elliptic-curve arithmetic circuit achieved a 17× gate reduction. The search ran for roughly 53 hours, largely autonomously. Human involvement was limited to the initial task specification. Occasional prompts to broaden the model's search direction were the only other interventions.
The team included Andi Gu and Alek Kemeny at Anthropic, and Susanne Yelin and Mikhail Lukin at Harvard University and QuEra Computing. The preprint is arXiv 2610.22367. If you're designing evaluation loops for optimization tasks, the paper's scoring function used a single Toffoli-gate objective — a form the model reduced incrementally across iterations.
---
For operators
Anthropic privately circulates draft AI safety bill to US lawmakers
Yahoo News (citing Politico) International Business Times SG CryptoBriefing
If state-level AI regulation is part of your current compliance planning, Anthropic's draft bill names what federal preemption of that landscape would look like. Politico reported October 17, citing three people with knowledge, that Anthropic representatives privately circulated draft legislation to Capitol Hill. As of October 17, the draft had not been formally introduced as legislation.
The draft would allow the federal government to block a frontier model's release only on a finding of "imminent enumerated risk." That threshold requires more than $1 billion in damages or more than 50 people killed or seriously injured. The eligible triggers are narrow: biological weapons creation, a model evading its developer's controls, or a cyberattack. A federal judge would have to order any halt.
Developer obligations in the draft: transparency reports, whistleblower protections, and critical safety incident reports to the Commerce Department. Independent evaluators registered with Commerce would be required; they must notify Commerce within 72 hours of identifying an imminent risk. Evaluators may be paid by the lab being evaluated, provided payment is not contingent on evaluation outcomes. The draft would preempt state laws on AI risk disclosures, third-party evaluations, and critical-incident reporting. If your compliance planning currently factors in state-level AI regulation, the preemption scope — risk disclosures, third-party evaluations, and critical-incident reporting — is what this draft would federalize.
---
Brave Leo routes AI chat through verifiable TEE enclaves, separating prompt content from sender identity
If you've been accepting a privacy policy as the practical ceiling for AI chat query privacy — where the guarantee is that the vendor says they won't log your prompts — Brave's October 15 announcement describes an architecture where that claim is technically verifiable. The Leo AI chat feature now routes user prompts through OHTTP into TEE-backed hardware. The relay sees who sent a request but not its contents; the compute node sees the contents but not who sent them.
The backend runs on Intel TDX CPU enclaves and Nvidia confidential-computing GPUs. The compute is operated by NEAR AI. The model is GLM 5.3 Flash. Z.ai released it as open-weight under MIT license in August 2026. Users can verify the TEE attestation against audited source code via Sigstore. Builds are reproducible against the published source. Every pipeline stage publishes its source. Contributors named in the Brave post: Brendan Eich, Ali Shahin Shamsabadi, and Darnell Andries.
Per Brave's blog post, the design makes the privacy claim structurally enforceable rather than policy-dependent: the relay cannot reconstruct which user sent which prompt, and users can independently verify that the code running in the enclave matches the published source. If you're building an AI-powered product and your privacy guarantee currently rests on a provider's policy rather than a technical constraint, evaluating whether OHTTP+TEE architecture fits your requirements is now a concrete option rather than a theoretical one.
---