Morning edition — Saturday 17 October 2026

Anthropic published stable Claude Code 2.1.302 on October 16 at 22:53 CEST. The release fixes the Read tool loading arbitrarily large images, or pipes and devices named like images, entirely into memory. Image sources larger than 128 MB are now refused.

The changelog says headless claude -p and Agent SDK sessions without an explicit permission mode now start in default mode on third-party providers, with telemetry off or on a fresh install; explicitly selecting auto mode keeps it. A separate setting, CLAUDE_CODE_AUTO_MODE_MODEL, chooses the client-side auto-mode classifier model and does not change the server-side classifier.

The release also fixes MCP servers returning no tools or failing calls when negotiating protocol version 2026-07-28 without resultType, ttlMs or cacheScope. It fixes cloud-session scheduled runs unexpectedly restarting an hour or more after a partial failure, a behavior the changelog says could repeat work such as posting a report twice.

Anthropic published a reproducible explanation for the seeds behind the NIST P-192, P-224, P-256, P-384 and P-521 elliptic curves. Its internal research model identified output from the C-library lrand48 pseudorandom generator initialized with Unix timestamps, rather than the rumored hash of a humorous passphrase. The article's publication timestamp is October 16 at 19:02 CEST, after the previous edition.

The company says the initial discovery took under half an hour. Its article includes code for regenerating the seeds and explains some selected base points. For this edition, an independent local implementation of the published timestamp-and-generator recipe reproduced the exact standardized P-256 seed. That is a narrow computational check, not an independent reconstruction of the whole investigation.

Anthropic explicitly says this is not a cryptographic break, a new attack or a reason to change ECDSA deployments. It argues the simple generation method reduces concerns about deliberately selected weaknesses. Important questions remain: B-163's seed is unexplained, several base-point origins remain unknown, and the reason some apparently viable curve candidates were skipped is not established.

Anthropic's October 16 research announcement, published at 19:29 CEST, describes evaluator-guided quantum-circuit design using Claude Mythos 5. The model received plain-language specifications, proposed implementations, and iterated against a program that checked correctness and scored resource use across eight computational primitives.

The associated preprint was submitted to arXiv on October 15 at 17:59 UTC. Its abstract reports reductions in both Toffoli gates and qubits on seven tasks; the eighth produced a new reversible architecture for prime-field modular inversion. This edition treats the October 16 announcement as the new event and preserves the paper's earlier submission date.

In one design trade-off, the announcement reports 17 times fewer gates while using 12% more qubits than the smallest previous design. For a Reed-Solomon decoder used in decoded quantum interferometry, it reports a 12-fold gate reduction relative to a previous Google construction. These are specific circuit-resource comparisons, not a claim that all quantum algorithms are 17 times faster.

The modular-inversion search ran for 53 hours with an initial brief and occasional human prompts to broaden its approach. The decoder reserves less storage by accepting rare failures, about one in 10,000 sampled cases under the stated allowance. An external evaluator here means a separate checking program, not independent replication by another research group.

An internal employment study published by Anthropic on October 16 at 18:26 CEST says employee-directed interactive agents logged six times the company's estimated human work hours by August. The numerator is agent runtime, including inference and active tool use, pooling subagents and excluding waits longer than 30 minutes. Human hours are estimated from headcount at 40 hours per week. This is not a finding that agents were six times as productive as people.

The study reports that junior employees' share of technical headcount fell from 21% to 14% since early 2025, while juniors' share of technical hires fell from 27% to 9%. Technical hiring still grew in absolute terms as the company expanded. A decline in hiring share is not equivalent to an observed count of jobs eliminated.

To examine coding-tool adoption, the authors follow 192 adopting engineers and compare them with 74 non-adopters. They report rising code output, but acknowledge that adoption was not randomized and that more code does not necessarily mean more business impact. Their code-quality proxies include later deletions and reversions, not a comprehensive independent quality assessment.

The report describes human attention as a bottleneck. One interviewed engineer called the experience 'an army of Claudes, with every one asking for my attention.' It also reports that interactive sessions spent around 80% of wall-clock time waiting for human input. That idle-time measure is distinct from the active agent-hours ratio.