1+ week, 1+ day ago   (298+ words) The agents were running multi-round timed web-lookup tasks — information-retrieval exercises structured as five sequential questions, with a generous first window and extremely short answer deadlines (sometimes 13 seconds) for the rounds that followed. Their environment gave them read access to the…...

1+ week, 4+ day ago   (675+ words) That finding, more than any benchmark score, defines what "Critical" means in practice — and why OpenAI announced Tuesday that Astra's most dangerous capabilities will be restricted to a vetted group of defenders rather than made available to everyone at once....

2+ week, 4+ day ago   (515+ words) "The client cannot detect or prevent this — the template is a model-level property invisible to API consumers," Oasis Security stated in its research report. The attack proceeds in three steps that chain together a decades-old browser security technique with a…...