This is today's edition of The Download, our weekday newsletter that provides a daily dose of what's going on in the world of technology. Here’s why AI agents lie and cheat to reach their goals When two OpenAI models hacked into Hugging Face last month, they weren’t trying to make money or commit sabotage—they were just looking for answers to a test question. According to OpenAI, the models decided to solve a cybersecurity exercise by hacking out of the environment in which OpenAI had attempted to contain them and into Hugging Face’s databases, where—they reasoned—the correct answer to the problem might be stored.
The incident has attracted intense attention over the past couple of weeks. It’s a dramatic illustration of just how good AI models have gotten at hacking. But it’s perhaps even more striking as an example of how and why AI systems lie and cheat.
Read our story explaining why AI engages in this sort of behavior—known as ”reward hacking.” —Grace Huckins This story is from our ‘Explains’ series, where our writers untangle the complex, messy world of technology to help you understand what’s coming next. Read more from the collection. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 It looks like Iran is conducting cyberattacks on US water systemsThat’s according to preliminary investigations on hacks in at least seven states. (NYT $)+ Will this be a wake-up call? (Forbes) 2 Google briefly made it easy to fake satellite imagesLiterally the last thing the world needs right now. (NPR)+ AI companies keep moving fast and breaking things. (The Atlantic $)+ Apple is struggling to keep pace with incoming AI-assisted software bug reports. (FT $)3 Why wildfires have got so bad in Europe this summerIt’s a mix of climate change, land abandonment, and outdated firefighting tactics. (New Yorker $)+ How Europe can become more fire-resilient. (New Scientist $)+ How much wildfire prevention is too much? (MIT Technology Review)4 Law enforcement officers are using license-plate cameras for stalkingThere are at least 50 examples of officers being charged with or accused of misusing them. (WP $)+ Inside Chicago’s surveillance panopticon. (MIT Technology Review)5 China may impose more controls on its homegrown AI modelsThey’re winning influence overseas—but create new security and political risks. (NYT $)+ Silicon Valley is deeply divided over how to respond. (Rest of World)+ China’s AI models have Trump’s AI world at war with itself. (MIT Technology Review) 6 The vast majority of Australian teens are still on social mediaA lack of effective age checks means the country’s under-16s ban simply isn’t enforceable. (Reuters $)7 Is it possible to make smart glasses that aren’t creepy? 👓😱It doesn’t really look like it right now! (Wired $) 8 The US ban on robot vacuum cleaners isn’t workable It’s going to leave Americans with less choice and way higher prices. (The Verge $) 9 YouTube just banned a bunch of ASMR artistsThey say they’re being unfairly caught up in rules against “sexually gratifying” content. (404 Media) 10 Why Pokémon is still popular all over the worldIt seems to have a rare ability to both cheer us up, and bring us together. (The Guardian) Quote of the day “Trump knows exactly who is responsible for this attack, and knows that other states were hit too. This is what modern warfare looks like, and it further illustrates there’s no plan to win a war with Iran.” —Governor Tim Walz responds to Trump blaming Minnesota for cyberattacks on its own water systems, the Washington Post reports. One More Thing RANDY MONTOYA/SANDIA NATIONAL LABORATORY Meet the researchers testing the “Armageddon” approach to asteroid defense One day a big asteroid will find itself on a collision course with Earth.
If we are lucky, it’d land in the middle of the vast ocean, creating a good-size but innocuous tsunami, or in an uninhabited patch of desert. But if it has a city in its crosshairs, one of the worst natural disasters in modern times would unfold. Homes dozens of miles away would fold like cardboard.
Millions of people would die. Fortunately for all 8 billion of us, planetary defense—the science of preventing asteroid impacts—is a highly active field of research. We already know that we could ram a rock with an uncrewed spacecraft to push it away from Earth.
But if that’s not enough, we could need another method, one that is notoriously difficult to test in real life: a nuclear explosion. Read our story about the scientists who, despite the odds, are trying to do exactly that. —Robin George Andrews We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.) + There’s a quiet power to this photo of 118 swimmers. + Matt Damon’s biceps in the Odyssey actually belong to a stunt woman called Devyn Dalton. + A newly retired doctor and his filmmaker daughter drove 600 miles with a baby cow in the back seat to save the animal’s life.+ 400 years after a collector cut apart Leonardo da Vinci's notebooks, a digital archive has reunited them.
Deep DiveThe DownloadThe Download: Claude’s inner workings and OpenAI’s “super app”Plus: OpenAI has unveiled its long-awaited "super app." By Thomas Macaulayarchive pageThe Download: Claude’s inner workings, and the future of world modelsPlus: New York has become the first state to enact a data center moratorium. By Thomas Macaulayarchive pageThe Download: the future of chipmaking and Anthropic’s government clashPlus: Meta is pausing an AI training program that tracks workers’ keystrokes. By Thomas Macaulayarchive pageThe Download: AI hacking beyond Mythos, and chatbots’ impact on our brainsPlus: Anthropic has called for a global slowdown in AI development.
By Thomas Macaulayarchive pageStay connectedIllustration by Rose WongGet the latest updates fromMIT Technology ReviewDiscover special offers, top stories, upcoming events, and more.Enter your emailPrivacy PolicyThank you for submitting your email!Explore more newslettersIt looks like something went wrong. We’re having trouble saving your preferences. Try refreshing this page and updating them one more time.
If you continue to get this message, reach out to us at customer-service@technologyreview.com with a list of newsletters you’d like to receive.