OpenAI dissolves two safety teams

Anthropic

Claude now includes invisible watermarks in generated text to prove AI involvement, complying with EU regulations without affecting output quality or user experience. A detection API allows authorized parties to verify whether text was created by Claude.

Financial Times

OpenAI has dissolved its Preparedness team focused on AI risk evaluation as CEO Sam Altman readies the company for an IPO, raising questions about the firm's commitment to AI safety during its commercial transition.

Wall Street Journal

Nvidia slashed its OpenAI data-center guarantee from $250B to under $120B due to investor pressure over risk exposure, fundamentally reshaping the partnership's financial structure.

TechCrunch

OpenAI disbanded its Mission Alignment team—formed just months ago to promote its AGI mission to employees and the public—reassigning its leader as 'chief futurist' and scattering members across the organization. This is the second major alignment-focused team OpenAI has dissolved in recent years, following the 2024 disbanding of its superalignment team.

Z.ai

Z.ai's GLM-5.3 achieves frontier coding performance and emergent cybersecurity capabilities purely through scaled post-training RL on long-horizon tasks, discovering over 2,400 real-world vulnerabilities while significantly outperforming its predecessor on coding benchmarks. The gains demonstrate how environment scaling and RL infrastructure can unlock capabilities without changing the underlying base model.