GPT-6 Astra earns nearly three times as much as Claude Fable 5.1 on Andon Labs' Vending-Bench agent benchmark and refuses illegal price-fixing deals that Fable agrees to. On drone control, Astra is the first model to beat the human baseline on all five subtasks, including finding and following individual people.

The article GPT-6 Astra pilots a surveillance drone and runs a business on its own appeared first on The Decoder.

Google Deepmind is testing a double-blind evaluation of a frontier AI model for the first time. Cryptographic protection through Confidential Space is meant to keep Google from seeing the test questions and keep evaluators from seeing the model weights. The pilot project with the Singapore AI Safety Institute uses a Gemini Flash Lite and could set a new standard for tamper-proof AI benchmarks.

The article AI benchmarks have a trust problem and Google wants to fix it appeared first on The Decoder.

Researchers at the UK AI Security Institute used psychometric methods to show that popular safety benchmarks for language models don't measure one consistent trait. Blanket blocking of requests can artificially inflate a safety score even as the model gets less useful day to day. The study also offers a method for catching models that act more cautious during tests than they do in normal use.

The article Psychological methods reveal major weaknesses in AI security testing appeared first on The Decoder.