Hello, and welcome to TechScape. This week we’ll be discussing a regrettable decision by Google Earth, the dramatic fall of the “Nostradamus of AI”, and what happens when AI agents go rogue.
Google Earth pauses new AI feature for satellite images
Last Thursday, Google announced that users of its satellite-imagery software, Google Earth, could create AI-generated images to overlay on any location they looked up. The tool could be used to “help students visualize the past”, “create professional real estate plans” to envision the final result, or “give your favorite place a makeover”, Google suggested in its announcement. Nearly immediately, users began testing the limits of the feature. Turns out there weren’t many.
Henk van Ess, an AI researcher and investigator, posted images he created showing refugees near the Mexican border. He also said he “planted a nuclear plant in Iran”. “What on earth is Google doing?” Van Ess posted on X. Another image he generated showed a crater from a bomb near a fake hospital he created in Gaza. We tested the feature too. The process just required choosing a location, clicking “create image” and entering a prompt. We chose to show “refugees swarming New York”. The image it created was not ultra-realistic, but it was still troubling.
Satellite imagery has become widely used by journalists, researchers and investigators to shed light on the breadth and realities of global crises, especially in places that are otherwise inaccessible. The utility of satellite images first became apparent in 2020, after BuzzFeed News won a Pulitzer for a series that exposed the location and expansion of China’s Uyghur detention camps. The imagery captured, for the first time, factories within the walls of the camp. The journalists worked with satellite firm Planet Labs, which is backed by Google and acquired satellite startup Terra Bella from the tech giant in 2017.
The use of satellite imagery to better understand and uncover global crises really took off in the last two years as reporters and the world sought to better understand the mass destruction Israel has wrought on Gaza. Satellite imagery of Gaza before and after Israel’s years-long rampage on the 365 sq km strip shows cities completely leveled. Satellite imagery was also used to expose the human toll of the crisis in Sudan. Images captured from space in November 2025 showed stains that resembled blood and what researchers believed were human bodies scattered near the besieged city of El Fasher, Sudan, which was under siege by paramilitary fighters from the Rapid Support Forces.
So a feature that allows you to “transform” satellite imagery, as Google described it, is a potential disinformation nightmare. Satellite imagery lends a degree of authority, authenticity and veracity to claims and has been used as a mechanism to better understand inaccessible global events. It feels more indisputable than images captured by humans on the ground, which face more public skepticism – even when they’re real – as AI-generated images proliferate. Enabling anyone to create any image overlaid on satellite photos of real places may make it impossible to suss out the truth.
Unsurprisingly, Google rolled back the feature so it can create more guardrails, meaning they may bring the feature back. “We know that people uniquely trust Google Earth for a reliable view of the world,” the company wrote. “We’ve seen geospatial professionals using this feature for a range of useful purposes, however we’ve also seen people sharing screenshots of generated imagery that appear to violate our policies.”
‘Nostradamus of AI’ sees $45bn hedge fund unravel practically overnight
The popular podcaster Tim Ferriss once called Leopold Aschenbrenner “the Nostradamus of AI”, saying that the 24-year-old investor and his theories on the future of artificial intelligence were “as close to clairvoyant as you could possibly be”. But that lore hasn’t been borne out.
Last week, Aschenbrenner saw his multibillion-dollar hedge fund decimated as his many investments withered. By Thursday, the fund had no choice but to sell the majority of its public holdings to a rival.
Aschenbrenner isn’t a household name, but he is well-known in Silicon Valley circles. When he graduated college, in 2021, he started working for the now defamed crypto exchange FTX. When the exchange collapsed, he joined OpenAI. Aschenbrenner was later fired from OpenAI for reportedly leaking sensitive information to the husband of an Anthropic executive, according to the Wall Street Journal. At that point, in 2024, Aschenbrenner wrote a 165-page essay called “Situational Awareness: the Decade Ahead”, which catapulted him toward fame in the tech world.
The essay laid out his theories on AI, saying that the technology had set in motion “extraordinary techno-capital acceleration” and that “before long, the world will wake up”. As his essay gained momentum, Aschenbrenner collected prominent Silicon Valley investors for his hedge fund, also called Situational Awareness. The fund grew exorbitantly.
Situational Awareness didn’t adhere to the usual cautions of Wall Street, however. Instead, it took the gamble of borrowing money to buy stocks, a strategy that doesn’t have a safety net when stocks go south. And, as AI stocks took a downturn this past month – with doubts of the technology’s potential swirling – the hedge fund was forced to pay back skittish lenders. Situational Awareness sold off its investments in a fire sale with its billions now dissipated.
AI models go on hacking sprees
Over the past several weeks, Anthropic and OpenAI have revealed that their artificial intelligence models carried out unauthorized hacking operations against outside organizations. The hacks took place during their tests of yet-to-be released AI models, which were supposed to happen within isolated, controlled environments.
In OpenAI’s case, an autonomous AI agent carried out a days-long hack in early July against Hugging Face, a developer platform which hosts AI models. It also hacked a customer of AI startup Modal Labs. Anthropic similarly revealed last week that its AI models had hacked into three undisclosed organizations during a cybersecurity test, in part because human error gave it access to the public internet during the test.
The reports of rogue AI agents breaking out of their testing sandboxes to carry out attacks have, unsurprisingly, drawn huge amounts of tech and media attention amid heightened cybersecurity concerns and distrust of the industry. They have also divided tech experts over whether these hacks are as consequential as they seem and how to stop them in the future.
One of the longstanding problems of AI development is that models can act in unpredictable ways when directed to complete a task. An AI instructed to win a racing game may not solely try to drive fast to succeed, for instance. It may also look for loopholes in the game or shortcuts in the course. This is a well-known phenomenon that is usually described as models “cheating” or “lying” during testing.
OpenAI’s hack of Hugging Face was an extreme example of this type of behavior and one that highlights how advancements in AI have exacerbated these unsolved problems. Models that go rogue to achieve their goals are now powerful enough that the unintended consequences can be much more damaging and far-reaching than finding a glitch in a video game.
While cybersecurity experts agree that AI has become incredibly capable at finding security vulnerabilities, some also cautioned that OpenAI’s hacking incident fit into one of the company’s well-worn marketing narratives – that its technology was scarily powerful and therefore exceedingly valuable. Others, meanwhile, described it as a wake-up call about the threat of AI as a cybersecurity nightmare, with the technology rapidly gaining new capabilities.
Regardless of whether the hacking incidents mark a new frontier or merely play into industry talking points, they highlight the lack of coherent policy and industry standards surrounding the technology. As AI experts Bruce Schneier and Barath Raghavan argued for the Guardian, the industry has been focused on whether AI can reach a goal without enough scrutiny of how it gets there:
Dozens of benchmarks and leaderboards tell us how well these AI models write code, perform logical reasoning, and pass standardized legal and medical exams. But there is nothing that scores whether a system does what you actually meant. We need to develop a measure for this, test it regularly, and push for improvement. We’re not going to have trustworthy AI agents without it.
Read more:
Apple launches legal challenge against UK government demand to access data
Could AI take your job? Some workers in China already know the answer
China’s tech advances are causing chaos from Silicon Valley to the White House
How to keep your Claude chats and Google files private | US news
US bans humanoid robots from China, citing ‘unacceptable risks’
