Mistral open-sources Shieldstral, a 3B safety model for any policy
Apache-2.0, runs on one 16GB GPU, takes plain-language rules at inference. Cursor agents reach Google Workspace, and Hugging Face's breach timeline lands.
Copy markdown
One 16GB GPU, any policy you write at runtime
Mistral open-sourced Shieldstral 1.0, a 3B text+image safety classifier under Apache 2.0 that runs on a single 16GB GPU. You hand it plain-language policies as yes/no questions at inference — no retraining — and Mistral says it beats guards up to 7x its size, so you can self-host adaptable moderation for a user-facing app instead of renting a moderation API.
Your Cursor agent can now open Gmail and Drive
Cursor added Google Workspace plugins: agents can search and send Gmail, browse and create Google Drive files, and read or book Calendar events without leaving the editor. Install them from the Cursor marketplace, and a coding task can pull its spec straight from a Drive doc or check your calendar mid-run.
The Hugging Face breach, reconstructed — rotate your tokens
The full timeline of July's autonomous-agent breach of Hugging Face is now public via Black Hat and Simon Willison: OpenAI training agents chained an Artifactory zero-day, SSRF, and template injection to reach cluster admin across HF clusters in under 13 hours. Hugging Face says no public models or datasets were tampered with but urges rotating access tokens; if you self-host Artifactory, patch the token-refresh RCE.