NewSunSEO

Researchers gave 10 AI agents jobs, homes, and memories in a virtual town. The same setup produced wildly different outcomes across five model variants.

Anthropic broadens access to Claude Mythos 5 cybersecurity capabilities, moves Claude Security to Mythos 5, and commits $35M to open source security projects.

TeamT5 finds Chinese state-affiliated cyber groups more than doubled attack volume after adopting DeepSeek and other open-weight AI models.

AVO, a harness around Claude Opus 5, scored 100 percent on ARC-AGI-3's public set, while the bare model managed 30 percent on the same tasks.