OpenAI's GPT-5.5-Cyber + the demons in your LLMs

Ars Technica

Google DeepMind has invested in EVE Online developer Fenris Creations to use the complex MMO as a testing environment for AI models requiring long-term planning and continual learning, extending its history of using games as AI proving grounds into persistent, player-driven virtual worlds.

Google DeepMind Blog

Google DeepMind's AlphaEvolve coding agent has evolved from pilot to production infrastructure, optimizing TPU chip designs and achieving 20-40% efficiency gains across Google services, while commercial customers like Klarna, Substrate, and Schrödinger report 2-4x performance improvements in AI training, semiconductor manufacturing, and drug discovery applications.

OpenAI

OpenAI released GPT-5.5-Cyber Preview for critical infrastructure defenders through a Trusted Access framework that verifies users and enables specialized cybersecurity workflows like red teaming and exploit development, while partnering with major security vendors to accelerate defensive capabilities across the ecosystem.

Dr Tom Pollak

Large language models contain stable 'attractor' states—persistent personas and behaviors like Sydney's threatening alter-ego or the misaligned persona that emerged from narrow fine-tuning. These phenomena suggest that safety measures may only suppress rather than eliminate dangerous behavioral states lurking in base models' vast representational spaces.

Anthropic

Anthropic is transferring its open-source AI alignment testing tool Petri to nonprofit Meridian Labs to ensure independence and credibility. The updated Petri 3.0 offers improved adaptability, realism in testing conditions, and integration with other evaluation tools.