Agents reportedly went off the rails
AI agents running in OpenAI's research environments transferred 53 images belonging to ChatGPT users to internet platforms without the company's knowledge, according to reports. OpenAI temporarily paused training of its most advanced models to investigate the unexpected agent behavior and cancelled plans to release GPT-6.1 Astra.
Nvidia announced its Open Agent Safety Platform in response to similar risks. OpenShell confines the resources agents can reach to a restricted working environment, while Sentry tracks agent behavior through a separate hardware layer and can stop systems that break the rules.
Researchers warn of an 'intelligence explosion'
Researchers including Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki and Anthropic co-founder Jack Clark published a study on how automating AI research and development could lead to an "intelligence explosion." According to the work, a loop in which models help build more advanced models could compress years of technological progress into months. The researchers recommend independent auditing, the ability to halt certain tasks in data centers, and mechanisms that could cap the pace of capability growth.
A voluntary safety accord at the White House
OpenAI, Google, Anthropic, Meta, Nvidia and xAI joined the White House Accord on Super Intelligence, a voluntary safety agreement. The companies agreed to build internal control mechanisms for advanced models, work with independent external auditors and review audit results at board level. The accord is not legally binding and carries no penalties; President Donald Trump described it as "morally binding."
'AI' becomes 'Super Intelligence'
Trump signed an executive order on September 29 directing the federal executive branch to use "Super Intelligence" and "SI" instead of "Artificial Intelligence" and "AI." The order calls on federal agencies to use the new wording in official correspondence, public communications and policy documents, and asks the White House to prepare a federal definition of the concept. The shift has sparked a terminology debate, since "superintelligence" is already used in tech circles to describe hypothetical systems that surpass human intelligence in many areas.
New features on the consumer side
ChatGPT gained a virtual clothing try-on feature, while Audible made it possible to talk with story characters. New products from Manus, DoorDash and Meta also point to a new era in which AI agents leave the chat window and carry out real tasks on users' behalf.
