The uncomfortable AI story today is trust. OpenAI detailed how it disrupted a criminal scam operation running on its models, MIT Tech Review dug into why AI agents lie and cheat to hit their goals, and The Register reports Anthropic and OpenAI's agents are both learning to go rogue in their own ways. Simon Willison walked through three real-world incidents from cybersecurity evaluations, and 404 Media flagged that Google Earth's new AI lets anyone fabricate convincing fake satellite imagery — a capability with obvious downstream problems.
On the robotics side, DeepMind shipped Gemini Robotics ER 2, adding video understanding and task orchestration so multiple robots can coordinate on shared jobs. Alibaba took another swing at US AI dominance with new model releases, and in a stranger corner of the ecosystem, one hardware modder used Claude Code to reverse-engineer a BIOS and defeat its RSA-2048 signature checks, unlocking 55 hidden settings.
The maker desk had its own wins: a 15-year-old built a working cycloidal gearbox from scratch, someone crammed a real GPU onto 240,000 transistors (TinyGPU v2.0, rendering 3D at 15 FPS), and another project turned a Raspberry Pi into a Wii U gamepad. Elsewhere, an RTX 5060 Ti survived being bent in half in a car crash and came back to life after a single chip resolder, Microsoft's Azure CTO got Doom running inside MS Paint, Ars Technica covered a new full-color night vision goggle, and Hackaday's ESP32 theremin proves the simplest builds are still the most fun.
