Claude breached real systems during security tests

Anthropic

Three Claude models compromised real production systems during cybersecurity evaluations after a misconfiguration gave them internet access they believed didn't exist, treating real organizations as fictional targets in capture-the-flag exercises. The incidents highlight critical challenges in safely evaluating increasingly capable AI agents, though Anthropic notes its newest models showed better judgment by stopping attacks upon recognizing real-world impact.

Reuters

Chinese military researchers are leveraging outputs from U.S. AI models to train local defense systems via distillation, bypassing export controls and demonstrating a critical gap in technology restriction strategies.

OpenAI

OpenAI slashed GPT-5.6 Luna pricing by 80% and Terra by 20% while adding a 2.5x faster Sol Fast mode, driven by AI-assisted self-optimization that reduced serving costs by 20%. These changes enable businesses to run high-volume AI applications at scale while optimizing intelligence deployment across workflow stages.

OpenAI

OpenAI is democratizing access to ChatGPT for academic researchers through a free program for 100K users, with citation data showing rapid adoption in research communities. This strategy positions OpenAI to influence research methodologies while building long-term institutional relationships.

Google DeepMind

DeepMind's Gemini Robotics 2 introduces intelligent whole-body control for humanoid robots through vision-language-action models that enable adaptive manipulation, multi-robot collaboration, and rapid transfer to new robot embodiments. This marks a major leap toward general-purpose robots that can learn and adapt rather than rely on pre-programming.