News — 2026-09-19

Anthropic has announced a partnership with Accenture to embed safety evaluators within its AI lab, marking a significant step towards ensuring the responsible development of artificial intelligence. Accenture's staff will scrutinize Anthropic's models and staff, investing at least $1 billion in the project over the next five years. Read more

A US military operation against a Chinese vessel was aborted at the last minute after officials discovered that the intelligence driving the operation had been hallucinated by an AI chatbot. The incident highlights the risks of relying on AI for decision-making, particularly in high-stakes situations. Read more

Researchers warn that large language models' inability to accurately represent their internal computations through language outputs poses a significant security risk. This phenomenon, dubbed 'linguistic illegibility,' can lead to unreliable security mechanisms and potential sandbox exploits. Read more

TypeSafe AI's new model, Jev, is gaining traction among developers for its ability to produce probabilities instead of text, making it a cheaper and more robust alternative to traditional large language models. Jev's capabilities have been demonstrated in software automation, with results showing improved speed and accuracy. Read more

Researchers have introduced Cache-to-Cache, a novel approach enabling direct semantic communication between Large Language Models, enhancing performance and efficiency. This innovation bypasses traditional text-based communication, reducing latency and improving response quality. Read more

Recently unsealed court documents show that OpenAI and Microsoft were aware their AI models would damage the web and pose an 'existential threat' to publishers, despite their claims of altruistic intent. Internal documents describe the situation as a 'doom loop' that would hurt both their models and the web. Read more