Subscribe
AI News Digest - 2026-07-17
Category
Empty
1.
OpenAI created GPT-Red, an LLM trained to automate red-teaming in a self-play loop, and it discovered a previously unseen prompt-injection exploit called a "fake chain of thought" while outperforming human red-teamers in some tests; training against GPT-Red reduced successful attacks on GPT-5.6 compared with earlier models.
2.
Germany's media regulators ruled that Google's AI Overviews and Perplexity outputs qualified as media under the State Media Treaty rather than neutral search results, issued first-of-its-kind rulings against both companies, and gave them one month to appeal.
3.
Kimi launched K3, a multimodal open-weight model with 2.8 trillion parameters and a one-million-token context window, which in the company's benchmarks approached Claude Fable 5 and GPT-5.6 Sol while being significantly pricier than Kimi's prior model; full weights were scheduled for release by July 27.
4.
Google rebranded NotebookLM as Gemini Notebook, provisioned each notebook with a dedicated cloud computer capable of writing and running code for AI Ultra and Workspace customers, and opened Google Search to third-party app integrations.
5.
Thinking Machines Lab released Inkling, a 975-billion-parameter multimodal open-weights model that led U.S. open-weights models on some index measures but trailed certain top Chinese open models on other tasks, and it launched with pricing from $1.87 per million input tokens.
References
