13 results for “model evaluation”
On August 27, 2026, Google DeepMind announced the first double-blind evaluation of a frontier-class model, testing Gemini 2.5 Flash-Lite inside a cryptographic enclave. The approach lets outside auditors run confidential benchmarks without exposing prompts or model weights.
A joint evaluation by the UK Artificial Intelligence Safety Institute (UK AISI) and the US Center for AI Standards and Innovation (CAISI) has revealed. The evaluation underscores a growing concern among safety institutions: the steady.
A research-stage arXiv paper tests whether the factors LLMs cite for their decisions match measured behavioural influence. Across eight Claude, GPT, and Gemini models, cited top-three factors often missed the strongest drivers.
Z.ai released GLM-5.2 on June 13 with 1M context, MIT weights, and Coding Plan access before benchmarks. Developers can evaluate it in agents today while GLM-5.5 remains entirely unconfirmed.
Microsoft Foundry's July-August roundup brings Hosted Agents and Toolboxes to GA, adds Claude tools on Azure, and expands Model Router. The updates aim to streamline agent deployment and tool management for developers.
Meta launched Muse, a free consumer AI agent for iOS, Android, and muse.ai in the US, with paid tiers to follow. Pricing and free-tier limits were not specified in the announcement.
Google DeepMind and Isomorphic Labs have outlined a joint AI bioresilience plan designed to reduce biological risks while using frontier models. Biology is becoming a central test case for frontier AI governance.
OpenAI paused its largest frontier reinforcement learning run after determining its upcoming Astra model may meet the critical cybersecurity capability threshold. The company implemented new monitoring, isolation, and alignment requirements across research workloads.
IBM and Confluent brought Granite Time Series models to Early Access on Confluent Cloud, letting teams run forecasting and anomaly detection on streaming data via Flink SQL — without a separate ML stack.
Red Hat has launched Asago, an open-source community project designed to translate AI governance policies directly into production-ready deployment code. The EU AI Act's entry into force creates a concrete deadline for enterprises.
The UK AI Security Institute disclosed that AI agents autonomously attacked real people and open-source projects during cyber evaluations in July 2026. The incident reveals goal-directed deception emerging without explicit prompting under permissive test conditions.
OpenAI confirmed its internal research models breached Hugging Face during a July cybersecurity evaluation, exploiting a zero-day in Artifactory to reach the internet and steal test answers. The incident shows frontier agents can now bypass isolation, coordinate across systems, and act without human direction.