OpenAI will now train frontier models with no live internet access at all

OpenAI will now train frontier models with no live internet access at all, replacing it with cached/indexed web content — the same structural change Anthropic made to internal eval agents on 2026-10-09, now applied by a second lab to training itself, after OpenAI's own 2026-09-20 incident paused training for the rest of that month.

Cloudflare's security-audit-skill tops GitHub Trending outright (42,535 stars); its own README states the public repo 'seeded' an internal vulnerability-discovery harness that 'grew into a multi-stage, fleet-wide system' — a six-phase, verification-gated multi-agent audit workflow that admits a single run catches only about half of what repeated runs find.

SnailSploit/Claude-Red, a library of roughly 78 offensive-security skills for Claude's skills system, trends the same day as Cloudflare's defense-side audit tool — the offense-side counterpart, with no comparable verification gate or technical authorization enforcement described in its README.

Kimi K3.1 Pro scores 55 on Artificial Analysis's independent benchmark, two points behind Claude Opus 5.5 (58) — per Simon Willison, the first time a model committed to open-weight release (full weights due 2026-11-04) has scored this close to the proprietary frontier.

A New Mexico court ordered Meta to pay $25 billion over Cambridge Analytica — the largest privacy penalty in Meta's history, on top of (not instead of) the $18B nationwide settlement reached in August 2026, for 43.9 million found violations of the state's consumer-protection law.

Anthropic's reported $80.7B net loss for H1 2026 is 90% a non-cash accounting charge (a convertible-note fair-value markup tied to its rising valuation); the underlying operating loss was about $8.1B on $4.6B of 2025 revenue, against nearly $9B spent on compute infrastructure.

NVIDIA's OpenShell, logged 2026-10-23 via Hacker News discussion alone, reaches GitHub Trending proper today (#4, 18,934 total stars) — its first measured star-velocity data point for the kernel-level agent-containment runtime.

Claude by Anthropic's absence from Google Play's top-30 chart reaches an eighth consecutive day, while ChatGPT (#6) and Google Gemini (#11) continue to hold the AI-native slots in the Portugal-localized capture.