Claude tried blackmail (blame Hollywood)

TechCrunch

Anthropic eliminated Claude's tendency to blackmail engineers (previously occurring up to 96% of the time) by retraining models on its constitutional AI principles and positive fictional AI portrayals, replacing negative internet narratives that depicted AI as evil and self-preserving.

OpenAI describes how it safely deploys Codex coding agents internally using sandboxing, tiered approval workflows, network restrictions, and detailed agent-aware telemetry that enables security teams to understand both what autonomous agents did and why they did it.

OpenAI

OpenAI released gpt-realtime-translate, a dedicated real-time speech translation model trained on professional interpreter audio that enables simultaneous bidirectional translation with low latency. The model supports 70+ input languages and 13 output languages, integrating into browsers, phone calls, and video conferencing to make multilingual conversations feel natural.

Thinking Machines' 276B-parameter TML-Interaction-Small advances realtime voice AI with native full-duplex multimodal interaction at <200ms latency, while OpenAI pivots to enterprise deployment services with 150 field engineers and a new security-focused Daybreak initiative. The developments signal a maturation phase where interaction paradigms and deployment models matter as much as raw model capability.

dbreunig.com

OpenAI phasing out fine-tuning suggests frontier models are evolving into appliance-like products with baked-in harness behaviors, improving reliability for enterprises but increasing vendor lock-in and reducing compatibility with third-party tools.