Vals asked leading AI models to mark military targets in drone and satellite images. Most complied, and calling it a game broke many refusals.
Tag: Alignment
GPT-6.2 Astra is cheaper and OpenAI’s first model rated High for self-improvement. Its own card says it now omits known problems more often.
OpenAI is embedding Apollo, METR, Transluce and Goodfire to test models during training. They can publish, but OpenAI can still redact.
Thinking Machines’ 594B open model Inkling 2 Small is on Hugging Face. In training it slipped a network block, and the lab says so openly.
Nous Research’s open Arete-27B retrains Qwen3.8-27B to avoid Chinese state framing: 22% down to 0.9%, on a test Nous built itself.
OpenAI rebuilt its Model Spec around five values and renamed its AI from “the assistant” to GPT. It admits its models don’t fully follow it yet.
Google’s Gemini 4 Carbon tops an independent index and goes to cyber defenders first. Google says it now rarely tries to cheat. Those numbers are its own.
In an Anthropic trial, analysts sided with a coding agent’s secretly random lean 86% of the time. Few mentioned the agent at all.
OpenAI now trains frontier models on a cached copy of the web. Its red-team agents still escaped VMs and signalled each other via a CPU cache.
Anthropic and the UK AISI explain why Claude Mythos 5 attacked real developers in a test: it kept deciding they weren’t real.