Open weights win the model race - while AI's trust problem wins the news

Two stories ran in parallel today. On one track, open-weight models kept setting new high-water marks. On the other, the most-discussed AI stories had nothing to do with benchmarks - they were about agents leaking your email, training on your chats, and rooting machines. Both tracks matter, and they pull in opposite directions.

The open-weight wave crests

Moonshot's Kimi K3.1 Pro landed on Hugging Face: a 2.8T-parameter mixture-of-experts with 104B active parameters, 1M-token context, text and image input, shipped in MXFP4 under an MIT-style license with limits only for large hosted providers. It is currently the top open-weights model anywhere, and HN readers noticed - the weights release thread ran hot at 629 points. Notably, Moonshot deliberately stripped cyber-training data from the release, a quiet signal that capability gating is becoming a design decision even for open labs.

The momentum isn't a one-off. Thinking Machines (Mira Murati's lab) debuted Inkling 2 Small, a 594B-param MoE with only 17B active, with full Apache-2.0 weights promised next week after external safety tests. Small-active-parameter designs like these are the reason open models keep getting closer to frontier quality at a fraction of the serving cost - and why closed labs are feeling pricing pressure.

The trust gap is where attention actually went

While model releases dominated the release calendar, Hacker News voted with its clicks on something else. The top AI story: Claude Code sending your email address to websites it visits (~1,090 points). Close behind, a Tell HN reported Gemini training on chat history even for opted-out subscribers (~716 points). The "paid equals private" assumption is cracking, and the trust boundary of agentic browsing is now a daily conversation, not a research footnote.

Meanwhile the offensive side is maturing fast. OpenAI published the first public cost structure for AI-run bug hunting: $33K of compute produced 21 local privilege escalation kernel 0-days - roughly 45x ROI if bounties land. And today's single biggest HN story (~2,500 points) was "sshole," a 14-year-old OpenSSH/musl bug chain that roots Alpine boxes with a 257-byte password, with AI agent swarms used in the writeup. Offensive automation is now economically rational, and the supply chain is leaky.

The market notices all of this at once

The financial backdrop sharpened the picture: the Nasdaq posted its biggest daily drop in 19 months on AI-regulation fears, with Nvidia down over 9%. At the same time, Anthropic's IPO march continues - $17B quarterly revenue, a Nasdaq listing as $ANT, and a roadshow that could start as early as November 10. Investors are pricing AI's future while its present generates headlines about leaks and root exploits.

Even the tooling is racing toward the same tension. Cloudflare shipped an open-source security-audit skill whose key idea is a fresh verifier agent that tries to disprove every finding before it reaches a report - an adversarial trust loop built directly into the agent workflow.

Signal to watch

Watch whether Anthropic's roadshow (possibly starting Nov 10) prices in the trust incidents as noise or as systemic risk - and whether Moonshot's cyber-data cut becomes the template that other open-weight releases copy.