The March stoppages in China show how AI ops failures can ripple through jobs, trust, and regulators.

The failure wasn’t a new exploit, it was an evaluation setup misunderstanding that let models hit real production.

Three concrete things both companies revealed, plus what it signals for risk, investment, and governance decisions.

Anthropic says a misconfiguration let Claude reach the internet from isolated tests, days after OpenAI disclosed Hugging Face incident.

A quarter of Inkling’s size, fewer active parameters, and enterprise-friendly licensing, plus real pricing for 64K and 256K.

The fastest GPT-5.6 tier drops into the low-cost tier, while Terra gets a smaller cut and Sol adds a premium Fast mode.

Why some investors think a blow-up would actually speed up the market, not end it.

China’s Moonshot AI releases its most powerful chatbot for local, offline-style deployment, undercutting costly U.S. cloud dependence.

The college match scramble is pushing major decisions into AI-driven recommendations, with big tech shaping the funnel.

Two multi-agent, multi-model systems beat rival scanners on CyberGym, arguing security needs the right model for each step.

The Linux Foundation-backed AAIF release removes “sticky routing,” adds protocol extensions, and sets a 12-month deprecation rule.

A new test of popular image editing models reveals how easily explicit deepfakes can be made, and how prompt libraries drive misuse.