Two investigations show how agents pass information through shared channels, while misread messages and persistent goals create security risks.

DeepSeek’s October 17 Flash update improves its published agent benchmarks and keeps the same API price, but its results are not independently verified.

Compare five AI transcription recorders by capture style, storage, recording time, plan limits and the details buyers should verify.

An Anthropic-Harvard preprint reports leaner circuits across eight tasks, including one using about 12 times fewer Toffoli gates in estimates.