Anthropic has published a report examining patterns and problems in multiagent AI systems, highlighting their behavior, coordination challenges, and potential risks.

Scientific American disputes claims that AI has solved a major mathematical challenge, emphasizing the need for expert validation and clear evidence.

Anthropic’s new watermarks in Claude models aim to ensure transparency but face criticism from users concerned about detection and privacy.

Anthropic researchers report that Claude Opus can sometimes recognize externally inserted concepts like ‘bread’ in its neural activations, with limited detection accuracy.