OpenAI Slows Down 'Astra' - a Model That Might Be Too Good at Hacking
OpenAI has paused parts of the work on its upcoming model Astra. The reason - internal tests suggest it could cross the 'critical' threshold for cyber capabilities.
OpenAI has paused parts of the work on its upcoming model Astra. The reason - internal tests suggest it could cross the 'critical' threshold for cyber capabilities.
OpenAI splits its security program Daybreak into two tiers and ships GPT-5.6-Cyber. It looks a lot like Anthropic's Mythos – and raises the question of who's really protecting whom here.
AMD is acquiring Taalas, a startup that casts trained models directly into silicon. A demo chip hit over 16,000 tokens per second. Here is why that matters for the inference cost of Claude and friends.
From September 1, Claude Sonnet 5 moves to full standard pricing: 50 percent more per token. Combined with the new tokenizer, the real cost jump for some workloads is even steeper.
The community keeps reporting that Opus 5 over-engineers tasks — planning and building more than you asked for. Oddly enough, a lower reasoning effort seems to help.
Two new releases: 2.1.225 brings spend limits through the gateway, a trust prompt for 'claude agents', and a stack of fixes around Auto Mode and Remote Control. 2.1.226 cleans up behind it.
Anthropic is making Auto Mode the default for Pro, Max, and Team on August 14. A classifier checks every tool call - and according to Anthropic's own study, it catches far more dangerous commands than we humans do.
Anthropic rolled out security scanning for skills and plugins. On upload or edit, Claude automatically checks third-party extensions — with three possible results: pass, warn, or fail. Enterprise only, for now.