OpenAI introduced GPT-Red, an internal red-team model that finds prompt-injection flaws at scale. Trained via self-play, it outperformed humans in attacks and has already reduced failure rates in newer GPT models through adversarial training.
Perplexity’s SPACE is a VM-based runtime for long-running AI agents, separating sessions from disposable sandboxes. It supports pause, resume, rollback, and recovery, while cutting sandbox startup times by 3.1 to 5x in production.
Google is working on a native menu for Skills on Gemini desktop. These Skills will potentially become available in all chats (and not only in Spark as it is now).
Users will be able to upload, create and edit their Gemini Skills, as well as use Gemini to create them.
Skill folders will be available there too, allowing users to select local folders containing necessary Skills.
MOONSHOT 🔥: Kimi K3 from Moonshot AI landed on the first place on Frontend Code Arena with 1679 points, outperforming Claude Fable 5 and GPT-5.6 Sol by far.
BREAKING 🔥: Kimi K3 is now available on the web and APIs! Both K3 Max and K3 Swarm Max options are available to users.
> Supports up to 1M context tokens. Optimized for coding, 3D gaming, and complex knowledge tasks. See documentation for details and usage instructions.
Moonshot AI appears close to launching Kimi K3, with leaks, teasers, and test sightings pointing to an imminent debut. Early reports suggest frontier-level performance, strong 3D and coding output, but slower runtimes on complex tasks.
