join.dev
AI news · 10 stories
OpenAI restricts GPT-6 Cyber to vetted security customers · Claude adds read-only personal finance connections · Google rolls out Gemini Omni 1.2 Flash
OpenAI’s November 5 system card describes GPT-6 Cyber, a specialized model for authorized vulnerability research and penetration testing. Its Critical cybersecurity designation comes with restricted Daybreak Red access, identity checks and organizational security requirements.
Read at OpenAI
Anthropic announced a US beta on November 5 that connects bank and card accounts through Plaid for spending analysis on Pro, Max and Ultra plans. Access is read-only, and conversations in the Personal Finance hub are excluded from model training, bringing account context into financial questions.
Read at Anthropic
Google announced Gemini Omni 1.2 Flash on November 5 with video generation and editing up to 15 seconds and multi-turn refinements. General availability across its API and business offering, alongside lower video pricing, gives creators more room to iterate on clips.
Read at Google
METR’s November 5 assessment estimates that Anthropic’s model capabilities on well-specified tasks are improving roughly twice as fast as before October 2025, with AI likely driving much of the acceleration. The evaluators emphasize substantial uncertainty and say Anthropic redacted much of the evidence, limiting public scrutiny.
Read at METR
Meta announced on November 5 that voice conversations and expressive avatars are rolling out in Muse on mobile, web and Mac, starting in the US. Users can interrupt responses and delegate background tasks while talking, expanding the assistant beyond text interaction.
Read at Meta
Baidu announced ERNIE 5.5 Flash Base on November 5, releasing a pretrained mixture-of-experts checkpoint with 280 billion total parameters and 14 billion active per token. The weights provide a starting point for domain adaptation and agent-training research under Baidu’s model license.
Read at Baidu
LeRobot v0.7.0, announced November 5, supports training directly from Hub datasets, Storage Buckets or LanceDB without downloading entire datasets first. Live policy interaction and remote inference let teams redirect robots while running larger models on separate GPU servers.
Read at Hugging Face
Box unveiled Architect on November 5 to combine frontier-model reasoning with enterprise content and existing procedures. The planned capabilities will design workflows, assemble execution tools and recommend improvements, extending content management into process automation.
Read at Box
Cloudflare launched personal-agent traffic analytics and granular access policies on November 4. Site owners can allow or block supported agents, including blocking them on ad-supported pages, giving publishers more control over automated visits made on users’ behalf.
Read at Cloudflare
Dharma-AI’s November 4 release describes a field extractor built on its two-billion-parameter OCR model with a task-specific adapter. It converts document images into typed JSON using open requests or supplied schemas, offering a specialized route from documents to structured business data.
Read at Dharma-AI