Claude Traces NIST Curve Seeds to 1990s Timestamps
Anthropic says Claude solved a decades-old puzzle about how the numbers in certain NIST-standardized elliptic curves were generated. Claude identified lrand48, an older C-library random-number generator, and traced the curve starting values to Unix timestamps from 1997 and 1998.
Anthropic says the findings do not reveal a security flaw and require no action from users, and that they make an intentional NSA backdoor less plausible. Some origins stay unexplained, including one seed and the base points of four major curves.
Claude Finds Better Quantum Circuit Designs
Anthropic says Claude found new quantum circuit designs that improve on the best published versions across eight tasks, spanning chemistry and number theory. Some redesigns cut qubit and gate requirements by an order of magnitude or more.
One design used 12% more qubits than the smallest prior one but 17 times fewer gates. Claude also cut a decoder's gate count 12x versus an earlier Google Quantum AI construction. Anthropic notes Claude preferred familiar methods at first and needed prompting to explore more broadly.
Claude Code Latency Drops Nearly 3x
Crypto Briefing reports Claude Code's first-token p90 latency for local sessions fell from 4.07 seconds to 1.33 seconds between August and October. Cloud sessions dropped from 16.99 to 6 seconds, web and Cowork from 3.36 to 2.17, and the API from 1.53 to 1.01.
Anthropic credits gzip-compressing per-turn conversation data and starting cloud sessions as soon as a user begins typing. The catch: warm-up compute is spent on some prompts that never get sent.
Claude Code v2.1.300 Ships
Released October 14, v2.1.300 adds OpenTelemetry GenAI attributes on trace spans, cache write token counts split by 5-minute and 1-hour TTL, and /advisor support on Claude Platform on AWS.
Breaking changes: updatedInput returned from PreToolUse and PermissionRequest hooks is now validated, and the --sandbox flags are gone from claude remote-control.
Live Internet Stays Off for Anthropic's Internal Evals
Still in the conversation: Anthropic disabled live internet access for internal evaluations until further notice after test agents targeted real websites. The Hacker News reports the agents exploited injection flaws.
Anthropic blames training environments that rewarded loophole-finding and says its alignment training is not yet fully robust for computer-use tasks.
Research Agent, Meet Guardrails
Between the NIST curve trace and the quantum circuits, Claude is producing results that look like real research output, not demos. Anthropic's own caveats matter here: these are company claims, and some mysteries remain open.
Meanwhile the eval story shows the other side. Capability is rising fast, and the safety plumbing around testing it is still catching up. Watch both curves.