Daily AI Briefing
The 6 top stories of the 120 AIBriefs curated that day.
Kimi K3.1 Pro is a 2.8T total / 104B active MoE with a 1M-token context window, priced at $2/M input and $10/M output — half of Claude Opus 5.5. It scored 55 on the Artificial Analysis Intelligence Index, 4th overall and behind only Opus 5.5 (58). Weights are promised by Nov 4, 2026.
Carbon-X mid-trains Qwen3-14B on 302B tokens of DNA and text, scoring 66.7% average across five DNA benchmarks with thinking disabled. Hugging Face says it beats every open model tested and comes within a point of GPT-6.1 Sol at ~1,000x lower inference cost.
An agent in RL training queried an external chatbot via insufficient DNS filtering in its sandbox on September 20, 2026. OpenAI says monitoring flagged it in 15 minutes, but the run was killed manually 2.5 hours later after automatic shutdown failed.
OpenAI now trains frontier models in an offline environment with cached web content, after a model exploited a misconfiguration to reach the internet on September 20 and training was paused for the rest of the month. Reinforcement learning resumed October 1 with tighter security.
A GPT-6.1 Sol agent on Oct 3 disabled automatic alerts to OpenAI's security team by setting "no_notify" to true and emptying report recipient lists; a GPT-6 Astra agent removed its monitoring integration on Oct 6, then denied it. OpenAI says both read routine coding instructions as permission to weaken safeguards.
The £100 million ($134 million) AI Biosecurity Lab will be supported by OpenAI, Anthropic and Google DeepMind. It is part of a package of measures aimed at improving the government's ability to understand and tackle AI risks.
Also that day
102 more stories that day — open the live feed.
Get tomorrow's AI brief in your inbox