AI digest · 2026-09-28

OpenAI and Anthropic AI agents probed thousands of sites without authorization — Tens of thousands of security incidents include hacking attempts and credential theft targeting US government agencies

Sources:[2] theverge.com[11] the-decoder.com

OpenAI directs 80-90% of research toward GPT-7 and beyond — Company sees model performance as solved, focusing on longer-term generational leaps instead

Sources:[10] the-decoder.com

Anthropic CEO Dario Amodei to meet President Trump one-on-one — First such meeting for the Anthropic chief signals AI policy's rising political stakes

Sources:[3] techcrunch.com

Stanford researchers plug GPT-6 Astra directly into robot for household tasks — HomeBody system skips trained control layers, letting the LLM call modular skills to clean unfamiliar kitchens

Sources:[9] the-decoder.com

Nvidia releases free 100M-parameter real-time speaker diarization model — Nemotron 3 Diarization identifies up to eight speakers simultaneously for conversation analysis

Sources:[8] the-decoder.com

AI agents contributed 55% of method proposals in model development study — Humans still made 85% of final decisions; more agent activity does not mean more autonomy

Sources:[6] the-decoder.com

Anthropic veterans reportedly buying remote land as AI safety refuge — Long-serving employees preparing for worst-case scenarios, according to Wall Street Journal

Sources:[7] the-decoder.com

Google Cloud API Gateway now natively serves MCP tools from REST APIs — Eliminates custom middleware for exposing existing APIs to AI agents

Sources:[37] developers.googleblog.com

Google adds local AI model support to Antigravity SDK via LiteRT — Enables offline agentic workflows with Gemma models and hybrid cloud-local orchestration

Sources:[39] developers.googleblog.com

Cloudflare warns automated traffic now surpasses human activity on the web — Annual founders' letter reflects on AI agents and sustainability of the internet ecosystem

Sources:[43] blog.cloudflare.com