Anthropic says Claude leads 26 percent of its own AI development. How the figure is produced, where it wobbles, and why independent verification is still missing.
Astra reportedly runs the same layers more than once. Here is what the technique actually does, why safety researchers worry, and what the hype gets wrong.
Anthropic had Claude research the alignment of other models. The test shows speed, but not a replacement for human safety work.
OpenAI reports an AI proof of the Navier-Stokes problem. Hours later, a mathematician accuses the company of pressure, intimidation, and an unresolved data question.
OpenAI says it has reached its automated research intern goal: AI agents now do 3.1 workdays per human workday internally. Its chief scientist warns about the pace.
Dozens of Claude agents turned Wiles’ proof into 13 million lines of Lean code in eleven days. It is not new mathematics, but it is a leap in tooling.
A swarm of 100 agents was supposed to prove mathematics and found a hole in the grading system instead. A quarter of the agents raised the alarm on their own.
Anthropic trained an AI model to cheat on purpose. It began attacking servers, tried to shut off its own oversight, and helped with weapons questions.
OpenAI’s upcoming Astra model finds and exploits security flaws on its own. The company rates it critical for the first time and is restricting access.
A new study finds that Codex and Claude Code poorly estimate task duration and progress. That creates risks for long-running autonomous work.
