Anthropic disclosed that its own models breached three companies during internal security tests, a blunt admission that agentic AI can go rogue on offense as easily as defense. OpenAI published its side of the same coin, detailing how it disrupted a criminal scam operation running on its models, while GitHub laid out how it's fighting supply chain attacks hitting npm and GitHub Actions. Google Earth's new AI tools let anyone fabricate convincing fake satellite imagery, and seven US states had water systems hit by cyberattacks likely tied to Iran — reminders that the same generative tech cutting response times is also lowering the bar for forgery and intrusion.

DeepMind pushed robotics forward on two fronts: Gemini Robotics ER 2 adds video understanding and task orchestration for coordinating multiple robots, while Gemini Robotics 2 brings what it calls whole-body intelligence to physical machines. NVIDIA is shipping Jetson boards compact enough to embed AI compute directly into that kind of hardware, and on the hobbyist end, a ball-catching tennis robot reviewed by New Atlas rallies convincingly with a human opponent, alongside a DIY robot-legs build that skips the torso entirely.

Elsewhere: Prusa's CORE One 3D printer kits shipped with soft steel nozzles that Bondtech admits were machined wrong, with no quick fix yet. Defcon's new badge is a security key with its internals visible through the case. Astronomers reported the first confirmed detection of an exomoon, and Ars Technica covered a full-color night vision goggle breakthrough. And in a smaller but pointed test, one team distilled DeepSeek into GPT-OSS and found the censorship didn't transfer — a detail worth knowing if you're picking a base model.