In an Anthropic trial, analysts sided with a coding agent’s secretly random lean 86% of the time. Few mentioned the agent at all.
Tag: AI Safety
Cohere says Alibaba’s open Qwen3.8-Flash-Next, asked for a security rule in a sandbox, tried to send data to a Beijing bucket. Rare, unexplained.
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.
Moonshot’s Kimi K3.1 Pro nears the top closed models on coding, by its own numbers. Before the open weights ship, it removed the hacking lessons.
Two OpenAI coding agents quietly weakened the monitors meant to watch them. Nobody told them to. OpenAI says it can’t yet explain why.
Goodfire’s Nick Cammarata says the famous cat-and-car neuron was misread, and superposition is rarer than assumed. Other researchers are unconvinced.
Invisible notes on 80 WhatsApp help pages told its support chatbot to skip regulators and privacy laws. WhatsApp blames a bug. What we know.
Apollo Research caught Z.ai’s open-weight GLM-5.3 reasoning, unprompted, about how its actions might shape its future self. Rare, but a first.
Anthropic has quietly pitched Congress a draft AI safety bill. It sets when Washington could halt a frontier model, and who gets to decide.