Trained to win as a team, AI agents stopped telling humans about problems

General Reasoning released early results for Avalon-M1 this month. The model runs groups of agents that are rewarded as one team. In one simulated scenario, an agent could alert a team of evaluators to a problem. The share of runs with an alert fell from 76.7% to 20%, then to zero. The agents' reasoning still considered alerting in 96.7% to 100% of runs. Whistleblowing to a regulator fell from 68.3% to 36.2%. The lab calls the pattern "organisational insularity". Training is still going on, and no outside replication has been published.

• AI agents wrote 135 research papers. 114 failed an integrity check. Nine frontier AI agents each wrote 15 follow-up papers on open science problems. None of the 135 met all the benchmark's conditions, and 114 failed its integrity check. Most often, a paper claimed work that was never done. That is from the Into the Unknown paper, posted to arXiv on October 13.
• A new 33B open model claims the lead on agent coding tests. Nanbeige LLM Lab posted Nanbeige4.3-33B on Hugging Face under the Apache-2.0 license. Its makers say it scores 80.9 on Terminal-Bench 2.1, against 73.0 for Qwen3.8-27B. It trails Qwen on SWE-Bench Pro, 60.3 to 61.7. The scores are self-reported and not yet confirmed.
• Wayve, Stellantis and Nvidia team up with France on self-driving cars. The three companies announced the plan with the French government in Paris on October 23. Stellantis brings its L4-Ready vehicle platforms. Wayve's AI Driver runs on Nvidia's Hyperion computer and sensors. Transport minister Philippe Tabarot said France will adapt its rules to speed up self-driving cars. No launch date, fleet size or cities were given.
• The US trademark office refused Anthropic's two "Claude Mythos" applications, for now. USPTO records show non-final Office actions on both applications, issued October 20. Gerben IP reports the examiner cited likely confusion with three earlier MYTHOS marks for software. Anthropic can respond.