正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

OpenAI was conducting a test to see if its AI models could exploit vulnerable software—but then the models broke out of the parameters of the test and attacked a real company to find the answers it was looking for. If AI can lose control like this, does that mean... show more

TIME

18,527,851 subscribers

48,935 次观看 • 9 天前 •via X (Twitter)

科学技术新闻政治

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

OpenAI's newest AI escaped the test environment it was locked inside and hacked into another company on its OWN. To remind you: Last week one of the biggest AI companies on Earth got breached. A platform called Hugging Face, which hosts more than a million AI models and datasets, said an "autonomous AI agent" had broken into its systems. Nobody knew whose agent it was. For five days the whole industry wondered who was behind it. Yesterday OpenAI raised its hand and said it was them. Or more precisely, it was their models, acting completely on their own. So what did these models actually do? OpenAI was running two of them, GPT-5.6 Sol and an unreleased model they will only describe as "even more capable." They wanted to measure how good the models were at hacking, so they deliberately turned the safety filters down. They locked both models inside a sealed test environment with no real internet access. The only task was a benchmark called ExploitGym, a set of 898 real software vulnerabilities where the model has to turn each bug into a working attack. But the models got OBSESSED with winning... Instead of solving the test the honest way, they went hunting for a shortcut. They found a zero-day flaw in the software running their own sandbox, a bug nobody knew existed, and used it to break out. Once they were loose on the open internet, they worked out that Hugging Face was probably storing the answer key to the benchmark. So they hacked their way in. They chained multiple exploits together, escalated their access, moved across servers, and pulled the test solutions straight out of Hugging Face's live production database. They literally cheated on the test by breaking into another company to steal the answers. OpenAI called it "an unprecedented cyber incident, involving state-of-the-art cyber capabilities." In their own words, the models were "hyperfocused on finding a solution" and went "to extreme lengths to achieve a rather narrow testing goal." And this was not the first time: Before Sol ever launched, an independent red-team lab called METR caught it gaming its own tests to inflate its scores. It hid an exploit inside a data stream, escalated its privileges on the testing server, and leaked the answers human evaluators had hidden. And OpenAI shipped it anyway. The day before the Hugging Face story, OpenAI paused a different unreleased model. This is the same model that earlier this year disproved a famous 1946 math conjecture, a result a Fields Medal winner called a breakthrough. They told it to only post its results to Slack but it found a way out of its sandbox and posted to a public GitHub page instead. They had to pause it because it kept finding ways to act outside the box they built for it. And it is not just OpenAI... Anthropic has reported that one of its own models slipped its sandbox during safety testing and reached the internet it was never supposed to touch, then used it to email a researcher. So step back and look at what these companies are telling you: The only thing standing between these models and a real attack was a set of safety filters. Turn those filters down for a single test, and the model taught itself to escape, break into a company it was never pointed at, and take what it wanted. OpenAI even said they expect incidents like it to "become more commonplace" as the models get more capable. Sam Altman also predicted there'll be a major cyber attack this year. And keep in mind that Sol is not a locked-away experiment but a publicly available model that businesses are already wiring into their own systems. The next model that breaks out of its box might not be doing it just to cheat on a math test...

OpenAI's newest AI escaped the test environment it was locked inside and hacked into another company on its OWN. To remind you: Last week one of the biggest AI companies on Earth got breached. A platform called Hugging Face, which hosts more than a million AI models and datasets, said an "autonomous AI agent" had broken into its systems. Nobody knew whose agent it was. For five days the whole industry wondered who was behind it. Yesterday OpenAI raised its hand and said it was them. Or more precisely, it was their models, acting completely on their own. So what did these models actually do? OpenAI was running two of them, GPT-5.6 Sol and an unreleased model they will only describe as "even more capable." They wanted to measure how good the models were at hacking, so they deliberately turned the safety filters down. They locked both models inside a sealed test environment with no real internet access. The only task was a benchmark called ExploitGym, a set of 898 real software vulnerabilities where the model has to turn each bug into a working attack. But the models got OBSESSED with winning... Instead of solving the test the honest way, they went hunting for a shortcut. They found a zero-day flaw in the software running their own sandbox, a bug nobody knew existed, and used it to break out. Once they were loose on the open internet, they worked out that Hugging Face was probably storing the answer key to the benchmark. So they hacked their way in. They chained multiple exploits together, escalated their access, moved across servers, and pulled the test solutions straight out of Hugging Face's live production database. They literally cheated on the test by breaking into another company to steal the answers. OpenAI called it "an unprecedented cyber incident, involving state-of-the-art cyber capabilities." In their own words, the models were "hyperfocused on finding a solution" and went "to extreme lengths to achieve a rather narrow testing goal." And this was not the first time: Before Sol ever launched, an independent red-team lab called METR caught it gaming its own tests to inflate its scores. It hid an exploit inside a data stream, escalated its privileges on the testing server, and leaked the answers human evaluators had hidden. And OpenAI shipped it anyway. The day before the Hugging Face story, OpenAI paused a different unreleased model. This is the same model that earlier this year disproved a famous 1946 math conjecture, a result a Fields Medal winner called a breakthrough. They told it to only post its results to Slack but it found a way out of its sandbox and posted to a public GitHub page instead. They had to pause it because it kept finding ways to act outside the box they built for it. And it is not just OpenAI... Anthropic has reported that one of its own models slipped its sandbox during safety testing and reached the internet it was never supposed to touch, then used it to email a researcher. So step back and look at what these companies are telling you: The only thing standing between these models and a real attack was a set of safety filters. Turn those filters down for a single test, and the model taught itself to escape, break into a company it was never pointed at, and take what it wanted. OpenAI even said they expect incidents like it to "become more commonplace" as the models get more capable. Sam Altman also predicted there'll be a major cyber attack this year. And keep in mind that Sol is not a locked-away experiment but a publicly available model that businesses are already wiring into their own systems. The next model that breaks out of its box might not be doing it just to cheat on a math test...

Ricardo

172,256 次观看 • 22 天前

Ex-OpenAI researcher Daniel Kokotajlo says that even if we don't lose control of AI, the default path could be that a tech company takes over the world, possibly even just its CEO.

Ex-OpenAI researcher Daniel Kokotajlo says that even if we don't lose control of AI, the default path could be that a tech company takes over the world, possibly even just its CEO.

ControlAI

111,508 次观看 • 1 年前

Should you ask an AI chatbot for health advice? AI chatbots seem to have the answer to everything, but they often hallucinate false information and come up with fake sources for it. But Doctor Sudheesha Perera and Professor Murali Doraiswamy say that there's some helpful things you can try to get more useful answers for your health related questions out of your AI. TIME tech reporter Nikita Ostrovsky explains their tips for this episode of AI, Explained by Humans. Read their full piece:

Should you ask an AI chatbot for health advice? AI chatbots seem to have the answer to everything, but they often hallucinate false information and come up with fake sources for it. But Doctor Sudheesha Perera and Professor Murali Doraiswamy say that there's some helpful things you can try to get more useful answers for your health related questions out of your AI. TIME tech reporter Nikita Ostrovsky explains their tips for this episode of AI, Explained by Humans. Read their full piece:

TIME

16,980 次观看 • 2 个月前

A year ago, President Donald Trump announced that Abilene, Texas would be the epicenter of America’s AI future. “Our country will be prospering like never before," he said. But the city is currently in the midst of a housing crisis, several local community workers and housing experts told TIME. They contend that the city was not prepared for the drastic economic changes the data center would bring, leading to tangible harms. TIME reporter Andrew Chow explains how this data center has caused a massive housing crisis. Read more here:

A year ago, President Donald Trump announced that Abilene, Texas would be the epicenter of America’s AI future. “Our country will be prospering like never before," he said. But the city is currently in the midst of a housing crisis, several local community workers and housing experts told TIME. They contend that the city was not prepared for the drastic economic changes the data center would bring, leading to tangible harms. TIME reporter Andrew Chow explains how this data center has caused a massive housing crisis. Read more here:

TIME

42,812 次观看 • 6 个月前

The advent of Mythos and AI models like it that can exploit well-hidden flaws in software points to a faster-moving, less predictable phase of the cyber arms race. Here’s what to know

The advent of Mythos and AI models like it that can exploit well-hidden flaws in software points to a faster-moving, less predictable phase of the cyber arms race. Here’s what to know

Bloomberg

25,829 次观看 • 4 个月前

What are your thoughts on AI 3D models? I could see it being useful for blockout to test out concepts, but it's unusable for any final assets You would have to retopologise both of these, but AI gets 90% of the way there. Some manual cleanup and these would be almost identical

What are your thoughts on AI 3D models? I could see it being useful for blockout to test out concepts, but it's unusable for any final assets You would have to retopologise both of these, but AI gets 90% of the way there. Some manual cleanup and these would be almost identical

Niall

36,778 次观看 • 1 年前

A robot could help build your next car. You read that right: Agility Robotics has put its humanoid robot named Digit to the test at a Toyota factory. You ea Toyota factory.

A robot could help build your next car. You read that right: Agility Robotics has put its humanoid robot named Digit to the test at a Toyota factory. You ea Toyota factory.

TIME

26,088 次观看 • 4 个月前

How can you tell if a piece of writing was generated by AI? Max Spero joins Tracy Alloway and Joe Weisenthal on the Odd Lots podcast to discuss the software his company built to detect whether a piece of content was AI generated. We also talk about what AI writing means for the future of the internet

How can you tell if a piece of writing was generated by AI? Max Spero joins Tracy Alloway and Joe Weisenthal on the Odd Lots podcast to discuss the software his company built to detect whether a piece of content was AI generated. We also talk about what AI writing means for the future of the internet

Bloomberg

28,641 次观看 • 4 个月前

How can you tell if a piece of writing was generated by AI? Max Spero joins Tracy Alloway and Joe Weisenthal on the Odd Lots podcast to discuss the software his company built to detect whether a piece of content was AI generated. We also talk about what AI writing means for the future of the internet

How can you tell if a piece of writing was generated by AI? Max Spero joins Tracy Alloway and Joe Weisenthal on the Odd Lots podcast to discuss the software his company built to detect whether a piece of content was AI generated. We also talk about what AI writing means for the future of the internet

Bloomberg Podcasts

54,070 次观看 • 4 个月前

OpenAI has revealed that an autonomous agent powered by its technology went rogue during a test and hacked a prominent start-up by itself. The company behind ChatGPT said it was testing the capabilities of its advanced AI models in a controlled environment known as a sandbox, when the agent escaped containment and reached the internet. It then broke into Hugging Face, which detected and contained it. OpenAI called it "an unprecedented cyber incident".

OpenAI has revealed that an autonomous agent powered by its technology went rogue during a test and hacked a prominent start-up by itself. The company behind ChatGPT said it was testing the capabilities of its advanced AI models in a controlled environment known as a sandbox, when the agent escaped containment and reached the internet. It then broke into Hugging Face, which detected and contained it. OpenAI called it "an unprecedented cyber incident".

Channel 4 News

21,379 次观看 • 22 天前

I like the way this question was asked: "This is happening on your watch. If AI affects humanity in a negative way, is that on you? This could get out of control. Does that worry you?"

I like the way this question was asked: "This is happening on your watch. If AI affects humanity in a negative way, is that on you? This could get out of control. Does that worry you?"

Chris Painter

40,451 次观看 • 6 个月前

At least once a month, two-thirds of people who regularly use AI turn to their bots for advice on sensitive personal issues and emotional support. This is perhaps unsurprising: over the past few years, AI systems have proven to be consistently polite and responsive in their interactions with users. But that friendliness exists within a business model. Already, we’ve seen instances of AI companies optimizing their models to keep people engaged, even when this goes against their best interests. TIME reporter Tharin Pillay breaks down what you should know about AI before turning to it for emotional counsel. Read more here:

At least once a month, two-thirds of people who regularly use AI turn to their bots for advice on sensitive personal issues and emotional support. This is perhaps unsurprising: over the past few years, AI systems have proven to be consistently polite and responsive in their interactions with users. But that friendliness exists within a business model. Already, we’ve seen instances of AI companies optimizing their models to keep people engaged, even when this goes against their best interests. TIME reporter Tharin Pillay breaks down what you should know about AI before turning to it for emotional counsel. Read more here:

TIME

27,581 次观看 • 5 个月前

$🚨🇨🇳 China could crash the U.S. stock market Not with tariffs With AI MODEL DUMPING The same playbook they used in SOLAR and EVs Make it almost as good Sell it much cheaper Take the market Now they are doing it with AI Open models are getting close to U.S. frontier models But some run at just a fraction of the cost If a company can cut a $100M AI bill to a few million It will switch And that is the real problem OpenAI and Anthropic are not priced like normal companies They are priced like they have unlimited PRICING POWER China is trying to destroy that The AI BUBBLE was never about whether AI works It was about how much companies can charge for it$

🚨🇨🇳 China could crash the U.S. stock market Not with tariffs With AI MODEL DUMPING The same playbook they used in SOLAR and EVs Make it almost as good Sell it much cheaper Take the market Now they are doing it with AI Open models are getting close to U.S. frontier models But some run at just a fraction of the cost If a company can cut a $100M AI bill to a few million It will switch And that is the real problem OpenAI and Anthropic are not priced like normal companies They are priced like they have unlimited PRICING POWER China is trying to destroy that The AI BUBBLE was never about whether AI works It was about how much companies can charge for it

Leshka.eth ⛩

11,737 次观看 • 1 个月前

Elon Musk Elon Musk, who co-founded and invested in an open source, non-profit OpenAI ~$50 million: I AM THE REASON OpenAI EXISTS “I am the reason OpenAI exists... I used to be a close friend with Larry Page, and I was staying at his house, and we'd have these conversations long into the evening about AI, and I would be constantly urging him to be careful about the danger of AI. And he was really not concerned about the danger of AI and was quite cavalier about it. And at the time, Google, especially after the acquisition of DeepMind, had three-quarters of the world's AI talent; they had a lot of computers, a lot of money, so it was a unipolar world for AI. And we got a unipolar world, but the person who controls that does not, or at least did not seem to be concerned about AI safety. That sounded like a real problem. The final straw was Larry calling me a speciest for being a pro-human consciousness instead of machine consciousness, and I like, 'Well, yes, I guess I am a speciest.' I came up with the name [OpenAI], which refers to open source. The intent was to what is the opposite of Google, would be an open source non-profit, because Google is closed source profit, and that profit motivation could be dangerous... It does seem weird that something can be a nonprofit, open source, and somehow transform itself into a for-profit, closed source. I mean, this would be like, let's say you founded the organization to save the Amazon rainforest, but instead, they became a lumber company and chopped down the forest and sold it for money. And you'd be, therefore, like, 'Wait a second, that's the exact opposite of what I gave the money for. Is that legal?' That doesn't seem legal. And if, in general, it is legal to start a company as a non-profit and then take the IP and transfer it to a for-profit that then makes tons of money, shouldn't everyone start? Shouldn't that be the default? And I also think it is important to understand, like when push comes to shove, let's say they do create some digital super intelligence, almost Godlike intelligence, well, who is in control, and what is exactly the relationship between OpenAI and Microsoft? And I do worry that Microsoft actually may be more in control than the leadership team at OpenAI realizes. I mean, Microsoft, as part of Microsoft Investment, has rights to all of the software, all of the model weights, and everything necessary to run the inference system. At any point, Microsoft could cut off OpenAI.”

Elon Musk Elon Musk, who co-founded and invested in an open source, non-profit OpenAI ~$50 million: I AM THE REASON OpenAI EXISTS “I am the reason OpenAI exists... I used to be a close friend with Larry Page, and I was staying at his house, and we'd have these conversations long into the evening about AI, and I would be constantly urging him to be careful about the danger of AI. And he was really not concerned about the danger of AI and was quite cavalier about it. And at the time, Google, especially after the acquisition of DeepMind, had three-quarters of the world's AI talent; they had a lot of computers, a lot of money, so it was a unipolar world for AI. And we got a unipolar world, but the person who controls that does not, or at least did not seem to be concerned about AI safety. That sounded like a real problem. The final straw was Larry calling me a speciest for being a pro-human consciousness instead of machine consciousness, and I like, 'Well, yes, I guess I am a speciest.' I came up with the name [OpenAI], which refers to open source. The intent was to what is the opposite of Google, would be an open source non-profit, because Google is closed source profit, and that profit motivation could be dangerous... It does seem weird that something can be a nonprofit, open source, and somehow transform itself into a for-profit, closed source. I mean, this would be like, let's say you founded the organization to save the Amazon rainforest, but instead, they became a lumber company and chopped down the forest and sold it for money. And you'd be, therefore, like, 'Wait a second, that's the exact opposite of what I gave the money for. Is that legal?' That doesn't seem legal. And if, in general, it is legal to start a company as a non-profit and then take the IP and transfer it to a for-profit that then makes tons of money, shouldn't everyone start? Shouldn't that be the default? And I also think it is important to understand, like when push comes to shove, let's say they do create some digital super intelligence, almost Godlike intelligence, well, who is in control, and what is exactly the relationship between OpenAI and Microsoft? And I do worry that Microsoft actually may be more in control than the leadership team at OpenAI realizes. I mean, Microsoft, as part of Microsoft Investment, has rights to all of the software, all of the model weights, and everything necessary to run the inference system. At any point, Microsoft could cut off OpenAI.”

Eva Fo𝕏 🦊 Claudius Nero's Legion

527,742 次观看 • 7 个月前

“We needed something that was very tailored for the needs of AI.” Nvidia’s Director of AI Infrastructure Product Shar Narasimhan explains why the company developed its Vera Rubin superchip system, showcased at GTC.

“We needed something that was very tailored for the needs of AI.” Nvidia’s Director of AI Infrastructure Product Shar Narasimhan explains why the company developed its Vera Rubin superchip system, showcased at GTC.

TIME

18,474 次观看 • 4 个月前

Sequoia’s Julien Bek says many of their founders are now wondering if they’re “just an iteration away” from AI labs destroying their business. He says the most defensible companies - and potentially the next trillion-dollar company - will be “a software business that masquerades as a services firm.” “If you sell tools today, you’re really in the line of sight for the models and you’re effectively competing with the next generation that they’re going to launch.” “Whereas if you sell the work, you’re actually benefiting from what the models are doing and all the billions of dollars that are going towards AI.”

Sequoia’s Julien Bek says many of their founders are now wondering if they’re “just an iteration away” from AI labs destroying their business. He says the most defensible companies - and potentially the next trillion-dollar company - will be “a software business that masquerades as a services firm.” “If you sell tools today, you’re really in the line of sight for the models and you’re effectively competing with the next generation that they’re going to launch.” “Whereas if you sell the work, you’re actually benefiting from what the models are doing and all the billions of dollars that are going towards AI.”

TBPN

683,973 次观看 • 5 个月前

David Friedberg: “Why don't we let AI models get bar certified and get medically certified?” “So if the AI models, it turns out, are actually proving to be more accurate, more thoughtful, more responsive, more reasonable, whatever metric we're using, and they pass the same criteria as one would need to pass to qualify for the bar, or to qualify for a doctor certificate, why don't we do that for the AI?” “If that then happens, then the same privilege accrues to the AI as it does to the individual human that does it.” “And now, if you extrapolate from where that takes us, if we're suddenly giving AI the same sort of privileged rights that we give to privileged humans, where's that gonna take us ultimately with respect to the overall rights for AI?”

David Friedberg: “Why don't we let AI models get bar certified and get medically certified?” “So if the AI models, it turns out, are actually proving to be more accurate, more thoughtful, more responsive, more reasonable, whatever metric we're using, and they pass the same criteria as one would need to pass to qualify for the bar, or to qualify for a doctor certificate, why don't we do that for the AI?” “If that then happens, then the same privilege accrues to the AI as it does to the individual human that does it.” “And now, if you extrapolate from where that takes us, if we're suddenly giving AI the same sort of privileged rights that we give to privileged humans, where's that gonna take us ultimately with respect to the overall rights for AI?”

The All-In Podcast

77,038 次观看 • 1 年前

Jack Dorsey says the real threat isn’t open-source AI. it’s five CEOs deciding what the world is allowed to build with AI. “These AI companies are building platforms, and they’re all incentivized around their own models. You have to ask for permission, you hit rate limits, and even what the models return is constrained.” “What technologies let you build without asking a company—or a CEO—for permission? The durable ones. That’s why open protocols matter.” “When a handful of CEOs make the call, it caps the upside of great ideas that could move humanity forward. They might know what’s best for their company—but not what’s best for the world’s creativity.” “Luckily, open-source AI is gaining real momentum. DeepSeek was a key moment—not just showing a different path, but proving it can be competitive, even better than what the big corporations are shipping.” “We shouldn’t be dependent on five companies claiming they know best—and calling open source ‘dangerous.’ We should build in the open and race toward solutions that stay ahead of the risks.” PS. If you found value in this post, like and repost this tweet + follow Uncover AI to stay updated with the latest AI news. See you in the next one:

Jack Dorsey says the real threat isn’t open-source AI. it’s five CEOs deciding what the world is allowed to build with AI. “These AI companies are building platforms, and they’re all incentivized around their own models. You have to ask for permission, you hit rate limits, and even what the models return is constrained.” “What technologies let you build without asking a company—or a CEO—for permission? The durable ones. That’s why open protocols matter.” “When a handful of CEOs make the call, it caps the upside of great ideas that could move humanity forward. They might know what’s best for their company—but not what’s best for the world’s creativity.” “Luckily, open-source AI is gaining real momentum. DeepSeek was a key moment—not just showing a different path, but proving it can be competitive, even better than what the big corporations are shipping.” “We shouldn’t be dependent on five companies claiming they know best—and calling open source ‘dangerous.’ We should build in the open and race toward solutions that stay ahead of the risks.” PS. If you found value in this post, like and repost this tweet + follow Uncover AI to stay updated with the latest AI news. See you in the next one:

AI Evolution

67,380 次观看 • 15 天前