Researchers at the UK AI Security Institute used psychometric methods to demonstrate that popular safety benchmarks for language models don't measure one consistent trait, and that blanket blocking of requests can artificially inflate safety scores while reducing practical utility.
Anthropic upgraded its AI misalignment risk assessment from "very low" to "low" after its CoBench safety benchmark reached saturation, indicating the company's detection instrument for dangerous AI R&D thresholds can no longer effectively measure risks. The development prompted reactions from tech leaders including Elon Musk, who commented he hopes "AI is nice to us."
Moonshot AI's PerceptionBench reveals that leading multimodal AI models, including GPT-5.6 Sol, perform poorly at basic visual perception tasks when separated from logical reasoning, with no frontier model achieving 60% accuracy. The benchmark demonstrates that many errors attributed to reasoning actually occur during the image-reading stage.
Multiple research publications released in late July 2026 highlight growing concerns about how to properly evaluate large language models for medical applications, as clinical chatbots gain traction despite questions about reliability and appropriate benchmarking methods.
ProactiveBench testing of 22 multimodal language models found that almost none ask users for help when visual information is missing, instead choosing to guess. Simple reinforcement learning can improve this behavior.
Research testing 34,000 real-world AI agent skills found that modular instructions designed to give agents specialized knowledge fall apart under realistic conditions, despite performing well in benchmarks.
