GhostApproval and Friendly Fire Expose the Trust Problem in AI Coding Agents
Two new attack patterns show how malicious repositories can turn coding agents, approval dialogs, and automated security reviews against developers.
Two new attack patterns show how malicious repositories can turn coding agents, approval dialogs, and automated security reviews against developers.

Cerebras used Gemma 4 to sample a rental-car video, detect visible damage, deduplicate findings, and annotate 17 frames in five seconds.

Mirelo and Kyutai released an open multi-instrument transcription model that turns mixed audio into editable, per-instrument MIDI.

OpenAI published a three-page AI-generated proof of a 50-year-old graph-theory conjecture. The manuscript is public, but expert review is not complete.

AI Studio apps can now claim a globally unique ai.studio subdomain, making prototypes easier to share while deployment still runs on Cloud Run.

The Long March 10B reached orbit on its maiden flight, then returned its first stage to a sea platform using a net-capture recovery system.

A former OpenAI researcher argues that superintelligence could create a dangerous concentration of compute, expertise, and political power.

Satya Nadella's AI strategy points to a durable advantage: owning the evals, feedback, context, and controls that make agents improve.

A developer-focused briefing on the week's most consequential AI releases, research findings, and production engineering signals.