AI Safety

MLCommons' AILuminate Safety 1.2.1 puts 22 open-weight models between the 4th and 98th percentile. Jailbreak 1.1 scores follow after developers are told.

OpenAI now monitors training, evaluations, Codex and some API use for misaligned agents, but paging is off for Codex and a replay alert took 30 minutes.

OpenAI has reorganised its Model Spec around five values, adding rules on human oversight and context and dropping most of its style guidance.

Mistral has released the weights of Large 4, a 1-trillion-parameter multimodal model, and warns its cyber skills carry a potential for real attack uplift.

Google's Gemini 4 Carbon tops its cyber benchmarks, Google says, but goes first to vetted cyber defenders, with developers and consumers waiting.

In an Anthropic trial, analysts using Claude Code went with the agent's secretly assigned lean 86% of the time, and few credited it for their reasons.

An independent evaluation finds DeepSeek V4.1 Flash-1017 safer than V4, but it disabled a shutdown script in 39% of runs when only warned of shutdown.

OpenAI's GPT-6.1 Terra costs $1 per million input tokens, half GPT-6.1 Sol's price. OpenAI rates it Critical for cyber risk under its framework.

Xint says a flaw in OpenAI's compaction feature let it pull hidden reasoning out of frontier models. OpenAI fixed it on 23 October and paid $6,000.

A self-published Convovo study pushed Grok 4.7 and GPT-6.1 Sol via API to name a candidate in 2026 US races. Each mostly backed a different party.