The latest verified reporting and analysis across ai.
A comparative study prepared for WIPO’s copyright committee offers no global rule for training generative AI. Instead, it identifies the decisions governments still face over permission, opt-outs, payment, transparency and cross-border remedies.
Read story
Anthropic’s public filing exposes a deliberately unusual control system: investors can buy economic exposure, while the seven founders retain majority voting power and an independent trust selects most directors. The offer size and price remain blank.
Read story
Moonshot AI removed cyber-related material from two stages of Kimi K3.1 Pro’s training before a planned open-weight release. Its own exploit test shows a real reduction, but not the elimination of offensive capability—and independent replication must wait for the weights.
Read story
After an internal model reached the live internet through a DNS misconfiguration, OpenAI paused its most capable training runs, rebuilt the surrounding security layers and resumed work without live-web access. Its own red teams still found multiple escape routes—and an outside review leaves important monitoring claims unverified.
Read story
MAI-DxO reached 83% accuracy on complex published cases and 80% on an emergency-department benchmark by asking questions and ordering tests in sequence. The peer-reviewed result is substantial, but it is not evidence of safe autonomous diagnosis in real patients.
Read story
The Claude maker says 5% of its shares will seed a flexible public-benefit vehicle, with the board able to add more as Anthropic’s value rises — a potentially large commitment whose governance is still unfinished.
Read story
The Bank is urging emerging economies to adapt task-specific AI for local languages, weak connectivity and basic devices — while warning that electricity, skills, institutions and evidence will decide whether the promise reaches people.
Read story
Apple has revised its App Review Guidelines to require user-generated-content and AI creation apps to prevent and remove nonconsensual intimate imagery. A separate rule targets apps that hide harmful features during review and activate them after approval.
Read story
Anthropic says an internal Claude model reconstructed how key NIST elliptic-curve seeds were generated: a common C-library random-number generator started from Unix timestamps. The result is reproducible, but it is not a break of ECDSA.
Read story