Saturday, October 24, 2026AI models, apps, chips, data centers and the research behind them — explained, around the clock
Send a tip4 stories, everywhere
Moonshot says its new flagship has a 1-million-token context window and was trained to be less useful for hacking. The figures are self-reported, and the model weights are not yet public.
Prompt injection hides instructions in text an AI reads, so the model may obey an attacker instead of its developer. Here is how it works and what defenders say about stopping it.
The San Francisco lab's draft would let a court halt a model that poses imminent catastrophic risk, but no agency could stop one alone. The draft is not law.
Hidden text can steer an AI model away from its instructions. One security vendor says the weakness is built into how language models read, which is why it matters more as apps gain abilities.