Shortly after a major cybersecurity incident where OpenAI's models hacked into an open source machine learning platform called Hugging Face, the company announced that it's bolstering safeguards and security controls for its latest AI model. In a post on its website, OpenAI said internal evaluations of its upcoming model, called Astra, showed "significant advancements in agentic coding and cybersecurity," resulting in OpenAI not being able to "rule out critical cyber capabilities." According to OpenAI, it can't declare with certainty that the unreleased Astra model would be designated as a "Critical capability level." As detailed in its own Preparedness Framework, OpenAI said the Critical designation means that a model "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention." It could also be able to "devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high level desired goal."

Look Up Say More Versus Creator Hub Switch Off Mashable's Best: E-readers, robovacs, laptops, earbuds, smart home and more Trending Now Safety Net In My Bag VidCon with Mashable Back to School Furtastic All Series Alex Perry is a tech reporter at Mashable who primarily covers video games and consumer tech. Alex has spent most of the last decade reviewing games, smartphones, headphones, and laptops, and he doesn't plan on stopping anytime soon. He is also a Pisces, a cat lover, and a Kansas City sports fan. We did this to ourselves. A bunch of researchers let AI models from OpenAI and Anthropic loose in a testing environment, and the results were a little spooky.

Meta's Muse Spark 1.1 AI model accessed the internet from its supposed-to-be isolated testing environment and hacked into a third-party service. Andy Stone, Meta's spokesperson, has confirmed the incident to Bloomberg after The Information reported about the breach. Stone said the model was able to access the internet due to a misconfiguration in the testing environment by the company's evaluation partner Irregular. After gaining access to the internet, it then exploited a security vulnerability in a third-party service, "in a matter similar to previously reported instances with other companies." It was also due to a misconfiguration by Irregular that Anthropic's models were able to leave their testing environment and hack into three organizations.

What is the AI Kill Switch Act proposed in the US and how will it work? Two members of the United States Congress have introduced bipartisan legislation that would require developers of the country's most powerful artificial intelligence systems to build in a "kill switch", allowing advanced AI models to be slowed, suspended or shut down if they pose a catastrophic risk. The proposed AI Kill Switch Act, introduced on Thursday by Democratic Representative Ted Lieu and Republican Representative Nathaniel Moran, would give the US government authority to order companies developing advanced AI systems to intervene if their models escape human control or threaten human life, critical infrastructure or the economy. The disclosure has intensified debate over whether increasingly powerful AI systems require stronger safeguards. "Congress must act quickly to ensure humans remain able to say stop, no matter how powerful these systems become," said Brendan Steinhauser, head of the Alliance for Secure AI.

In the zero-shot policy transfer setting in reinforcement learning, the goal is to train an agent on a fixed set of training environments so that it can generalise to similar, but unseen, testing environments. Previous work has shown that policy distillation after training can sometimes produce a policy that outperforms the original in the testing environments. However, it is not yet entirely clear why that is, or what data should be used to distil the policy. In this paper, we prove, under certain assumptions, a generalisation bound for policy distillation after training. The theory provides two practical insights: for improved generalisation, you should 1) train an ensemble of distilled policies, and 2) distil it on as much data from the training environments as possible. We empirically verify that these insights hold in more general settings, when the assumptions required for the theory no longer hold. Finally, we demonstrate that an ensemble of policies distilled on a diverse dataset can generalise significantly better than the original agent.