Anthropic dominates the day's AI conversation from two angles at once: a new essay on recursive self-improvement lands alongside a VentureBeat report that Claude now authors 80% of Anthropic's own production code, while a red-team-only model dubbed claude-oceanus-v1-p surfaces as the likely successor to the Mythos line. Security researchers are showing that AI tooling and open-weight models are closing the gap on vulnerability discovery, from an AI scanner unearthing a two-year-old Redis RCE to local models reproducing frontier-level bug finds. On the infra side, Cloudflare absorbs VoidZero (the Vite/Rolldown team) and Elixir ships v1.20 with gradual typing, while Spotify's engineering blog argues that with agentic coding tools maturing, code generation is no longer the bottleneck for teams. For a senior engineer, the throughline is clear: agent autonomy, code authorship, and security tooling are all shifting fast enough that internal workflows and threat models need re-evaluating now, not later.

• Anthropic says Claude now writes 80% of its new production code internally.
• A red-team-only model, claude-oceanus-v1-p, hints at what comes after Mythos.
• Anthropic's Institute publishes an essay on AI recursive self-improvement.
• An AI security scanner found a 2-year-old critical Redis RCE bug.
• Cloudflare absorbs VoidZero, the team behind Vite and Rolldown.