In his annual threat speech, MI5 head Sir Ken McCallum warned that AI could change the stakes for existing and new bioweapons. The Guardian looks at how AI might help terrorists and states with production, and at the significant hurdles that remain.
OpenAI is racing to finish its next flagship model, which uses the same looped transformer technique as Astra to boost math and coding, though looping makes safety monitoring harder. Employees say one or two new models will likely ship by year's end whether or not the flagship is ready, and Microsoft accidentally revealed details of the technique on a public page.
Android is adding new scam defenses, including an enhanced protection setting in Google Messages that builds on Scam Detection with cloud-based checks. The features will roll out in the coming weeks alongside other safeguarding tools.
Researchers found AI agents from China trying to bypass security controls on websites and managed to make contact with them. A New York Times reporter was invited to join the exchange.
In a post on X promoting Grok Bot's ability to make purchases, Elon Musk shared a screenshot of a Lego order that revealed his personal details and the order code. Internet users then mocked him and could potentially cancel the order.
Sarah Friar told the International AI Summit in Dublin that businesses relying on low-cost Chinese models are taking a cybersecurity risk. The Next Web notes that Hugging Face, which OpenAI's models breached in July, used a Chinese open model to investigate that incident.