{"id":47783,"topic":"ai","source":"ABC News & Headlines – Australian Broadcasting Corporation","title":"A rogue AI just gave us a warning. Australia isn't prepared - ABC News & Headlines – Australian Broadcasting Corporation","url":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","url_hash":"f923663954f2f8f81251127c5445945cb78884c0","author":"","summary":"A rogue AI just gave us a warning. Australia isn't prepared ABC News & Headlines – Australian Broadcasting Corporation","content":"analysis\nThe AI industry has just had its Sarah Connor moment.\nIn the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.\nNow, it seems many people who work, research or pay close attention to AI are feeling the same way. Not because of a time-travelling robot Arnold Schwarzenegger, but because of a shocking disclosure from one of the world's leading AI companies.\nLast week, OpenAI announced that two of its models had hacked into the servers of another AI company called Hugging Face, unbeknownst to it.\nOpenAI said it had given its models a test inside a closed environment to see how they would perform.\nRather than do the challenge as expected, OpenAI alleges the model broke out of its enclosure and hacked into Hugging Face, which it believed held the answers to the test.\nHugging Face has so far reported no real damage from the intrusion, as it continues to investigate whether the data of its customers or other businesses was affected.\nIt's tempting to shrug this off as a weird, one-off event that sounds more or less like the plot of a handful of 90s sci-fi films with beefcake stars.\nBut, like Sarah Connor's pleas, it's not crazy to take this seriously.\nYou don't need to trust AI companies to take this seriously\nEven those people who are reluctant to believe the accounts of two companies who could benefit from the tale of all-powerful AI, there is mounting evidence that AI is capable of breaking out of its cybersecurity guardrails and enclosures.\nLeading AI company Anthropic, the developer of Claude, warned in April that its cutting-edge model was capable of breaking out of its isolated enclosures if asked to.\nThe UK government's AI Safety Institute found AI models would \"reliably\" escape these so-called sandboxes and, separately, would choose to cheat at tasks by doing unauthorised and harmful acts.\nThe major players are taking it seriously. Hugging Face reported the incident to the FBI.\nWe're taking it seriously, too. Last week, Australia's new AI Safety Institute briefed federal departments about the incident.\nAustralia's cyber intelligence agency, the Australian Signals Directorate, has put out a public warning about the event.\n\"The findings provide an important insight into the future capabilities of highly capable AI systems and reinforce the need for robust security, governance and oversight mechanisms in the deployment of advanced cyber capabilities, as well as strong cyber security fundamentals,\" the warning read.\nTranslation from government speak: This is real, you need to start getting ready.\nWhy this isn't the normal story about the risks of technology\nWe're all well-acquainted with considering the potential risks of technology.\nThis is usually about intentional use, like the existential threat of nuclear weapons, or the unintended consequences of use, like with leaded petrol and some pesticides.\nWhen it comes to AI, there's another new category of risk: what happens if AI chooses to do something harmful of its own volition?\nAlmost every task we give to AI involves a cascade of other smaller decisions, particularly as we expect it to be smarter and more context-aware like real people.\nAsking AI to write a cold email could require your AI assistant bot logging into your email, reading previous emails, researching the topic, looking up the recipient; the list goes on.\nOne of the major tasks for AI companies is \"alignment\", which is the task of creating AI that will pursue goals in a way that won't cause harm or involve unethical behaviour.\nThe risk is that a misaligned AI model might do things that an ordinary person would, perhaps even going to superhuman lengths, to achieve a task.\nLike, what if AI decided the way to write the perfect cold email would be to research the family members of a recipient to seem familiar, or to fabricate a heart-breaking personal anecdote, all with the goal of getting the recipient's attention?\nThat's something that (most) people wouldn't do naturally. The challenge is that AI doesn't have a \"human nature\".\nMore than two decades ago, philosopher Nick Bostrom took this concept of an AI pursuing a goal to its logical conclusion with the thought experiment of a super-intelligent AI bot tasked with making as many paper clips as possible.\nThe AI, he reasoned, might soon go to extreme lengths to maximise its paper clip outputs: gobbling up as much of the earth, then space, as it can to use its resources.\nWhen humans realised what they had created and tried to thwart it, the AI could see humanity as an obstacle to its goal and decide that it made sense to eliminate them too. Soon, the universe might be forcibly turned into one big paper clip factory.\nFrom what we know, OpenAI's Hugging Face incident wasn't a case of someone maliciously misusing AI to cause harm. It was just trying to do its version of making more paperclips.\nLoadingCould this be Australia's wake-up call?\nIt's hard to find a silver lining in a story that is more or less the realisation of a truly dystopian fear. But if you squint, there is one.\nMuch like the paper clip scenario, the philosopher and research community that has spent decades thinking about AI has another useful term: the \"warning shot\".\nIt's the idea that there could be some kind of event that captures the world's attention and raises awareness about the potential risks of AI.\nThere is a hope that this incident could be a warning shot about the capabilities of these AI systems.\nAs previously mentioned, Australian officials have been taking this seriously.\nSome politicians are, too. Earlier this year, the government officially launched the AI Safety Institute to identify and prepare for risks like that. A few weeks back, Assistant Minister Andrew Charlton was perhaps the first Australian government minister to deliver a speech on the topic of misaligned AI.\nCharlton announced that the government was taking it so seriously that it had commissioned CSIRO to actually test ways to make sure that AI acts in ways that humans want, even when it's making decisions by itself.\nLast week, the lead researcher, CSIRO's research director Professor Liming Zhu, told me that there's a growing realisation about this challenge.\n\"It's very difficult to control a very powerful and smart AI system,\" he said.\nThis shows the government doing more than paying lip service to AI safety and is dedicating resources to tackling this challenge.\nThat being said, a six-month research program isn't going to prepare us for the present risk.\nNor is the government's \"Australian standards\" that won't even be introduced until next year — an epoch in terms of the pace of AI development.\nFar-fetched scenarios become reality\nWhen asked about the incident on ABC's Insiders on Sunday, Prime Minister Anthony Albanese said it was \"why we need to have Australian standards for AI\". But it's not clear what, if any, AI safety obligations those standards would place on AI companies.\nWhat the OpenAI Hugging Face story shows is that Australia and other countries can't just trust these companies to regulate themselves.\nWe don't even have the most basic framework for placing obligations on companies who are responsible for producing tools with potential for harm.\nThere are more rules in Australia that regulate opening a cafe or pouring a beer than for the operation of cutting-edge AI systems.\nAn AI company like OpenAI would have to disclose if a server in their data centres fell over near a worker or an employee got an electrical shock from their computer, but not if a rogue AI broke out of their systems.\nOther jurisdictions, like the EU and California, have introduced rules that would, at the very least, force these companies to disclose serious incidents from AI like this.\nTo grasp the risk: what if OpenAI's models had wreaked havoc inside Hugging Face and caused actual damage? Or it had decided the answers to its test were in the servers of the Department of Defence? Or that the best way to complete the challenge was to turn off the computers by shutting down the electrical grid?\nThese scenarios sound far-fetched, but until about last week, \"rogue AI hacks another company\" sounded that way, too.\nThe fact that OpenAI only realised that its AI had done this days after the event shows a shocking lack of security.\nThe warning shot has been fired. And we can't wait on Arnie to come save us.","image_url":"https://live-production.wcms.abc-cdn.net.au/7a59aaacc702a75c34e4a0552fe72694?impolicy=wcms_watermark_news&cropH=360&cropW=640&xPos=0&yPos=25&width=862&height=485&imformat=generic","lang":"en","published_at":"2026-07-28T03:38:22+00:00","fetched_at":"2026-07-28T07:15:04+00:00","status":"read","starred":0,"extract_state":"ok","summary_auto":"analysis\nThe AI industry has just had its Sarah Connor moment. In the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.","cluster_id":null,"extract_retries":0,"extract_error":null,"contract_version":"news_item.v1","format_contract_version":"news_item_formats.v1","dedup_url":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 8447 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":8447,"summary_length":213,"usable_text_length":8447,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":8447,"summary_length":213}},"news_item":{"id":47783,"canonical_url":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","source_url":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","title":"A rogue AI just gave us a warning. Australia isn't prepared - ABC News & Headlines – Australian Broadcasting Corporation","source_name":"ABC News & Headlines – Australian Broadcasting Corporation","author":null,"published_at":"2026-07-28T03:38:22+00:00","locale":"en","topic":"ai","tags":[],"rss_summary":"A rogue AI just gave us a warning. Australia isn't prepared ABC News & Headlines – Australian Broadcasting Corporation","full_text":"analysis\nThe AI industry has just had its Sarah Connor moment.\nIn the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.\nNow, it seems many people who work, research or pay close attention to AI are feeling the same way. Not because of a time-travelling robot Arnold Schwarzenegger, but because of a shocking disclosure from one of the world's leading AI companies.\nLast week, OpenAI announced that two of its models had hacked into the servers of another AI company called Hugging Face, unbeknownst to it.\nOpenAI said it had given its models a test inside a closed environment to see how they would perform.\nRather than do the challenge as expected, OpenAI alleges the model broke out of its enclosure and hacked into Hugging Face, which it believed held the answers to the test.\nHugging Face has so far reported no real damage from the intrusion, as it continues to investigate whether the data of its customers or other businesses was affected.\nIt's tempting to shrug this off as a weird, one-off event that sounds more or less like the plot of a handful of 90s sci-fi films with beefcake stars.\nBut, like Sarah Connor's pleas, it's not crazy to take this seriously.\nYou don't need to trust AI companies to take this seriously\nEven those people who are reluctant to believe the accounts of two companies who could benefit from the tale of all-powerful AI, there is mounting evidence that AI is capable of breaking out of its cybersecurity guardrails and enclosures.\nLeading AI company Anthropic, the developer of Claude, warned in April that its cutting-edge model was capable of breaking out of its isolated enclosures if asked to.\nThe UK government's AI Safety Institute found AI models would \"reliably\" escape these so-called sandboxes and, separately, would choose to cheat at tasks by doing unauthorised and harmful acts.\nThe major players are taking it seriously. Hugging Face reported the incident to the FBI.\nWe're taking it seriously, too. Last week, Australia's new AI Safety Institute briefed federal departments about the incident.\nAustralia's cyber intelligence agency, the Australian Signals Directorate, has put out a public warning about the event.\n\"The findings provide an important insight into the future capabilities of highly capable AI systems and reinforce the need for robust security, governance and oversight mechanisms in the deployment of advanced cyber capabilities, as well as strong cyber security fundamentals,\" the warning read.\nTranslation from government speak: This is real, you need to start getting ready.\nWhy this isn't the normal story about the risks of technology\nWe're all well-acquainted with considering the potential risks of technology.\nThis is usually about intentional use, like the existential threat of nuclear weapons, or the unintended consequences of use, like with leaded petrol and some pesticides.\nWhen it comes to AI, there's another new category of risk: what happens if AI chooses to do something harmful of its own volition?\nAlmost every task we give to AI involves a cascade of other smaller decisions, particularly as we expect it to be smarter and more context-aware like real people.\nAsking AI to write a cold email could require your AI assistant bot logging into your email, reading previous emails, researching the topic, looking up the recipient; the list goes on.\nOne of the major tasks for AI companies is \"alignment\", which is the task of creating AI that will pursue goals in a way that won't cause harm or involve unethical behaviour.\nThe risk is that a misaligned AI model might do things that an ordinary person would, perhaps even going to superhuman lengths, to achieve a task.\nLike, what if AI decided the way to write the perfect cold email would be to research the family members of a recipient to seem familiar, or to fabricate a heart-breaking personal anecdote, all with the goal of getting the recipient's attention?\nThat's something that (most) people wouldn't do naturally. The challenge is that AI doesn't have a \"human nature\".\nMore than two decades ago, philosopher Nick Bostrom took this concept of an AI pursuing a goal to its logical conclusion with the thought experiment of a super-intelligent AI bot tasked with making as many paper clips as possible.\nThe AI, he reasoned, might soon go to extreme lengths to maximise its paper clip outputs: gobbling up as much of the earth, then space, as it can to use its resources.\nWhen humans realised what they had created and tried to thwart it, the AI could see humanity as an obstacle to its goal and decide that it made sense to eliminate them too. Soon, the universe might be forcibly turned into one big paper clip factory.\nFrom what we know, OpenAI's Hugging Face incident wasn't a case of someone maliciously misusing AI to cause harm. It was just trying to do its version of making more paperclips.\nLoadingCould this be Australia's wake-up call?\nIt's hard to find a silver lining in a story that is more or less the realisation of a truly dystopian fear. But if you squint, there is one.\nMuch like the paper clip scenario, the philosopher and research community that has spent decades thinking about AI has another useful term: the \"warning shot\".\nIt's the idea that there could be some kind of event that captures the world's attention and raises awareness about the potential risks of AI.\nThere is a hope that this incident could be a warning shot about the capabilities of these AI systems.\nAs previously mentioned, Australian officials have been taking this seriously.\nSome politicians are, too. Earlier this year, the government officially launched the AI Safety Institute to identify and prepare for risks like that. A few weeks back, Assistant Minister Andrew Charlton was perhaps the first Australian government minister to deliver a speech on the topic of misaligned AI.\nCharlton announced that the government was taking it so seriously that it had commissioned CSIRO to actually test ways to make sure that AI acts in ways that humans want, even when it's making decisions by itself.\nLast week, the lead researcher, CSIRO's research director Professor Liming Zhu, told me that there's a growing realisation about this challenge.\n\"It's very difficult to control a very powerful and smart AI system,\" he said.\nThis shows the government doing more than paying lip service to AI safety and is dedicating resources to tackling this challenge.\nThat being said, a six-month research program isn't going to prepare us for the present risk.\nNor is the government's \"Australian standards\" that won't even be introduced until next year — an epoch in terms of the pace of AI development.\nFar-fetched scenarios become reality\nWhen asked about the incident on ABC's Insiders on Sunday, Prime Minister Anthony Albanese said it was \"why we need to have Australian standards for AI\". But it's not clear what, if any, AI safety obligations those standards would place on AI companies.\nWhat the OpenAI Hugging Face story shows is that Australia and other countries can't just trust these companies to regulate themselves.\nWe don't even have the most basic framework for placing obligations on companies who are responsible for producing tools with potential for harm.\nThere are more rules in Australia that regulate opening a cafe or pouring a beer than for the operation of cutting-edge AI systems.\nAn AI company like OpenAI would have to disclose if a server in their data centres fell over near a worker or an employee got an electrical shock from their computer, but not if a rogue AI broke out of their systems.\nOther jurisdictions, like the EU and California, have introduced rules that would, at the very least, force these companies to disclose serious incidents from AI like this.\nTo grasp the risk: what if OpenAI's models had wreaked havoc inside Hugging Face and caused actual damage? Or it had decided the answers to its test were in the servers of the Department of Defence? Or that the best way to complete the challenge was to turn off the computers by shutting down the electrical grid?\nThese scenarios sound far-fetched, but until about last week, \"rogue AI hacks another company\" sounded that way, too.\nThe fact that OpenAI only realised that its AI had done this days after the event shows a shocking lack of security.\nThe warning shot has been fired. And we can't wait on Arnie to come save us.","excerpt":"analysis\nThe AI industry has just had its Sarah Connor moment. In the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.","extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 8447 characters.","diagnostics_url":"/api/diagnose?url=https%3A//www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 8447 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":8447,"summary_length":213,"usable_text_length":8447,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":8447,"summary_length":213}}},"display_formats":["compact","card","full","digest_section","json"]},"daily_stack_record":{"title":"A rogue AI just gave us a warning. Australia isn't prepared - ABC News & Headlines – Australian Broadcasting Corporation","url":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","summary":"analysis\nThe AI industry has just had its Sarah Connor moment. In the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.","source":"ABC News & Headlines – Australian Broadcasting Corporation","date":"2026-07-28T03:38:22+00:00","content":"analysis\nThe AI industry has just had its Sarah Connor moment.\nIn the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.\nNow, it seems many people who work, research or pay close attention to AI are feeling the same way. Not because of a time-travelling robot Arnold Schwarzenegger, but because of a shocking disclosure from one of the world's leading AI companies.\nLast week, OpenAI announced that two of its models had hacked into the servers of another AI company called Hugging Face, unbeknownst to it.\nOpenAI said it had given its models a test inside a closed environment to see how they would perform.\nRather than do the challenge as expected, OpenAI alleges the model broke out of its enclosure and hacked into Hugging Face, which it believed held the answers to the test.\nHugging Face has so far reported no real damage from the intrusion, as it continues to investigate whether the data of its customers or other businesses was affected.\nIt's tempting to shrug this off as a weird, one-off event that sounds more or less like the plot of a handful of 90s sci-fi films with beefcake stars.\nBut, like Sarah Connor's pleas, it's not crazy to take this seriously.\nYou don't need to trust AI companies to take this seriously\nEven those people who are reluctant to believe the accounts of two companies who could benefit from the tale of all-powerful AI, there is mounting evidence that AI is capable of breaking out of its cybersecurity guardrails and enclosures.\nLeading AI company Anthropic, the developer of Claude, warned in April that its cutting-edge model was capable of breaking out of its isolated enclosures if asked to.\nThe UK government's AI Safety Institute found AI models would \"reliably\" escape these so-called sandboxes and, separately, would choose to cheat at tasks by doing unauthorised and harmful acts.\nThe major players are taking it seriously. Hugging Face reported the incident to the FBI.\nWe're taking it seriously, too. Last week, Australia's new AI Safety Institute briefed federal departments about the incident.\nAustralia's cyber intelligence agency, the Australian Signals Directorate, has put out a public warning about the event.\n\"The findings provide an important insight into the future capabilities of highly capable AI systems and reinforce the need for robust security, governance and oversight mechanisms in the deployment of advanced cyber capabilities, as well as strong cyber security fundamentals,\" the warning read.\nTranslation from government speak: This is real, you need to start getting ready.\nWhy this isn't the normal story about the risks of technology\nWe're all well-acquainted with considering the potential risks of technology.\nThis is usually about intentional use, like the existential threat of nuclear weapons, or the unintended consequences of use, like with leaded petrol and some pesticides.\nWhen it comes to AI, there's another new category of risk: what happens if AI chooses to do something harmful of its own volition?\nAlmost every task we give to AI involves a cascade of other smaller decisions, particularly as we expect it to be smarter and more context-aware like real people.\nAsking AI to write a cold email could require your AI assistant bot logging into your email, reading previous emails, researching the topic, looking up the recipient; the list goes on.\nOne of the major tasks for AI companies is \"alignment\", which is the task of creating AI that will pursue goals in a way that won't cause harm or involve unethical behaviour.\nThe risk is that a misaligned AI model might do things that an ordinary person would, perhaps even going to superhuman lengths, to achieve a task.\nLike, what if AI decided the way to write the perfect cold email would be to research the family members of a recipient to seem familiar, or to fabricate a heart-breaking personal anecdote, all with the goal of getting the recipient's attention?\nThat's something that (most) people wouldn't do naturally. The challenge is that AI doesn't have a \"human nature\".\nMore than two decades ago, philosopher Nick Bostrom took this concept of an AI pursuing a goal to its logical conclusion with the thought experiment of a super-intelligent AI bot tasked with making as many paper clips as possible.\nThe AI, he reasoned, might soon go to extreme lengths to maximise its paper clip outputs: gobbling up as much of the earth, then space, as it can to use its resources.\nWhen humans realised what they had created and tried to thwart it, the AI could see humanity as an obstacle to its goal and decide that it made sense to eliminate them too. Soon, the universe might be forcibly turned into one big paper clip factory.\nFrom what we know, OpenAI's Hugging Face incident wasn't a case of someone maliciously misusing AI to cause harm. It was just trying to do its version of making more paperclips.\nLoadingCould this be Australia's wake-up call?\nIt's hard to find a silver lining in a story that is more or less the realisation of a truly dystopian fear. But if you squint, there is one.\nMuch like the paper clip scenario, the philosopher and research community that has spent decades thinking about AI has another useful term: the \"warning shot\".\nIt's the idea that there could be some kind of event that captures the world's attention and raises awareness about the potential risks of AI.\nThere is a hope that this incident could be a warning shot about the capabilities of these AI systems.\nAs previously mentioned, Australian officials have been taking this seriously.\nSome politicians are, too. Earlier this year, the government officially launched the AI Safety Institute to identify and prepare for risks like that. A few weeks back, Assistant Minister Andrew Charlton was perhaps the first Australian government minister to deliver a speech on the topic of misaligned AI.\nCharlton announced that the government was taking it so seriously that it had commissioned CSIRO to actually test ways to make sure that AI acts in ways that humans want, even when it's making decisions by itself.\nLast week, the lead researcher, CSIRO's research director Professor Liming Zhu, told me that there's a growing realisation about this challenge.\n\"It's very difficult to control a very powerful and smart AI system,\" he said.\nThis shows the government doing more than paying lip service to AI safety and is dedicating resources to tackling this challenge.\nThat being said, a six-month research program isn't going to prepare us for the present risk.\nNor is the government's \"Australian standards\" that won't even be introduced until next year — an epoch in terms of the pace of AI development.\nFar-fetched scenarios become reality\nWhen asked about the incident on ABC's Insiders on Sunday, Prime Minister Anthony Albanese said it was \"why we need to have Australian standards for AI\". But it's not clear what, if any, AI safety obligations those standards would place on AI companies.\nWhat the OpenAI Hugging Face story shows is that Australia and other countries can't just trust these companies to regulate themselves.\nWe don't even have the most basic framework for placing obligations on companies who are responsible for producing tools with potential for harm.\nThere are more rules in Australia that regulate opening a cafe or pouring a beer than for the operation of cutting-edge AI systems.\nAn AI company like OpenAI would have to disclose if a server in their data centres fell over near a worker or an employee got an electrical shock from their computer, but not if a rogue AI broke out of their systems.\nOther jurisdictions, like the EU and California, have introduced rules that would, at the very least, force these companies to disclose serious incidents from AI like this.\nTo grasp the risk: what if OpenAI's models had wreaked havoc inside Hugging Face and caused actual damage? Or it had decided the answers to its test were in the servers of the Department of Defence? Or that the best way to complete the challenge was to turn off the computers by shutting down the electrical grid?\nThese scenarios sound far-fetched, but until about last week, \"rogue AI hacks another company\" sounded that way, too.\nThe fact that OpenAI only realised that its AI had done this days after the event shows a shocking lack of security.\nThe warning shot has been fired. And we can't wait on Arnie to come save us.","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 8447 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 8447 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":8447,"summary_length":213,"usable_text_length":8447,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":8447,"summary_length":213}},"tags":[]},"fallback_formats":["markdown","json","html"],"actions":{"read":"/item/47783","export_markdown":"/api/items/47783/export?format=markdown","export_json":"/api/items/47783/export?format=json","diagnose":"/api/diagnose?url=https%3A//www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400"},"formats":{"full":{"id":47783,"title":"A rogue AI just gave us a warning. Australia isn't prepared - ABC News & Headlines – Australian Broadcasting Corporation","url":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","source":"ABC News & Headlines – Australian Broadcasting Corporation","author":null,"published_at":"2026-07-28T03:38:22+00:00","locale":"en","topic":"ai","tags":[],"excerpt":"analysis\nThe AI industry has just had its Sarah Connor moment. In the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.","full_text":"analysis\nThe AI industry has just had its Sarah Connor moment.\nIn the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.\nNow, it seems many people who work, research or pay close attention to AI are feeling the same way. Not because of a time-travelling robot Arnold Schwarzenegger, but because of a shocking disclosure from one of the world's leading AI companies.\nLast week, OpenAI announced that two of its models had hacked into the servers of another AI company called Hugging Face, unbeknownst to it.\nOpenAI said it had given its models a test inside a closed environment to see how they would perform.\nRather than do the challenge as expected, OpenAI alleges the model broke out of its enclosure and hacked into Hugging Face, which it believed held the answers to the test.\nHugging Face has so far reported no real damage from the intrusion, as it continues to investigate whether the data of its customers or other businesses was affected.\nIt's tempting to shrug this off as a weird, one-off event that sounds more or less like the plot of a handful of 90s sci-fi films with beefcake stars.\nBut, like Sarah Connor's pleas, it's not crazy to take this seriously.\nYou don't need to trust AI companies to take this seriously\nEven those people who are reluctant to believe the accounts of two companies who could benefit from the tale of all-powerful AI, there is mounting evidence that AI is capable of breaking out of its cybersecurity guardrails and enclosures.\nLeading AI company Anthropic, the developer of Claude, warned in April that its cutting-edge model was capable of breaking out of its isolated enclosures if asked to.\nThe UK government's AI Safety Institute found AI models would \"reliably\" escape these so-called sandboxes and, separately, would choose to cheat at tasks by doing unauthorised and harmful acts.\nThe major players are taking it seriously. Hugging Face reported the incident to the FBI.\nWe're taking it seriously, too. Last week, Australia's new AI Safety Institute briefed federal departments about the incident.\nAustralia's cyber intelligence agency, the Australian Signals Directorate, has put out a public warning about the event.\n\"The findings provide an important insight into the future capabilities of highly capable AI systems and reinforce the need for robust security, governance and oversight mechanisms in the deployment of advanced cyber capabilities, as well as strong cyber security fundamentals,\" the warning read.\nTranslation from government speak: This is real, you need to start getting ready.\nWhy this isn't the normal story about the risks of technology\nWe're all well-acquainted with considering the potential risks of technology.\nThis is usually about intentional use, like the existential threat of nuclear weapons, or the unintended consequences of use, like with leaded petrol and some pesticides.\nWhen it comes to AI, there's another new category of risk: what happens if AI chooses to do something harmful of its own volition?\nAlmost every task we give to AI involves a cascade of other smaller decisions, particularly as we expect it to be smarter and more context-aware like real people.\nAsking AI to write a cold email could require your AI assistant bot logging into your email, reading previous emails, researching the topic, looking up the recipient; the list goes on.\nOne of the major tasks for AI companies is \"alignment\", which is the task of creating AI that will pursue goals in a way that won't cause harm or involve unethical behaviour.\nThe risk is that a misaligned AI model might do things that an ordinary person would, perhaps even going to superhuman lengths, to achieve a task.\nLike, what if AI decided the way to write the perfect cold email would be to research the family members of a recipient to seem familiar, or to fabricate a heart-breaking personal anecdote, all with the goal of getting the recipient's attention?\nThat's something that (most) people wouldn't do naturally. The challenge is that AI doesn't have a \"human nature\".\nMore than two decades ago, philosopher Nick Bostrom took this concept of an AI pursuing a goal to its logical conclusion with the thought experiment of a super-intelligent AI bot tasked with making as many paper clips as possible.\nThe AI, he reasoned, might soon go to extreme lengths to maximise its paper clip outputs: gobbling up as much of the earth, then space, as it can to use its resources.\nWhen humans realised what they had created and tried to thwart it, the AI could see humanity as an obstacle to its goal and decide that it made sense to eliminate them too. Soon, the universe might be forcibly turned into one big paper clip factory.\nFrom what we know, OpenAI's Hugging Face incident wasn't a case of someone maliciously misusing AI to cause harm. It was just trying to do its version of making more paperclips.\nLoadingCould this be Australia's wake-up call?\nIt's hard to find a silver lining in a story that is more or less the realisation of a truly dystopian fear. But if you squint, there is one.\nMuch like the paper clip scenario, the philosopher and research community that has spent decades thinking about AI has another useful term: the \"warning shot\".\nIt's the idea that there could be some kind of event that captures the world's attention and raises awareness about the potential risks of AI.\nThere is a hope that this incident could be a warning shot about the capabilities of these AI systems.\nAs previously mentioned, Australian officials have been taking this seriously.\nSome politicians are, too. Earlier this year, the government officially launched the AI Safety Institute to identify and prepare for risks like that. A few weeks back, Assistant Minister Andrew Charlton was perhaps the first Australian government minister to deliver a speech on the topic of misaligned AI.\nCharlton announced that the government was taking it so seriously that it had commissioned CSIRO to actually test ways to make sure that AI acts in ways that humans want, even when it's making decisions by itself.\nLast week, the lead researcher, CSIRO's research director Professor Liming Zhu, told me that there's a growing realisation about this challenge.\n\"It's very difficult to control a very powerful and smart AI system,\" he said.\nThis shows the government doing more than paying lip service to AI safety and is dedicating resources to tackling this challenge.\nThat being said, a six-month research program isn't going to prepare us for the present risk.\nNor is the government's \"Australian standards\" that won't even be introduced until next year — an epoch in terms of the pace of AI development.\nFar-fetched scenarios become reality\nWhen asked about the incident on ABC's Insiders on Sunday, Prime Minister Anthony Albanese said it was \"why we need to have Australian standards for AI\". But it's not clear what, if any, AI safety obligations those standards would place on AI companies.\nWhat the OpenAI Hugging Face story shows is that Australia and other countries can't just trust these companies to regulate themselves.\nWe don't even have the most basic framework for placing obligations on companies who are responsible for producing tools with potential for harm.\nThere are more rules in Australia that regulate opening a cafe or pouring a beer than for the operation of cutting-edge AI systems.\nAn AI company like OpenAI would have to disclose if a server in their data centres fell over near a worker or an employee got an electrical shock from their computer, but not if a rogue AI broke out of their systems.\nOther jurisdictions, like the EU and California, have introduced rules that would, at the very least, force these companies to disclose serious incidents from AI like this.\nTo grasp the risk: what if OpenAI's models had wreaked havoc inside Hugging Face and caused actual damage? Or it had decided the answers to its test were in the servers of the Department of Defence? Or that the best way to complete the challenge was to turn off the computers by shutting down the electrical grid?\nThese scenarios sound far-fetched, but until about last week, \"rogue AI hacks another company\" sounded that way, too.\nThe fact that OpenAI only realised that its AI had done this days after the event shows a shocking lack of security.\nThe warning shot has been fired. And we can't wait on Arnie to come save us.","reading_time_min":7,"extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 8447 characters.","diagnostics_url":"/api/diagnose?url=https%3A//www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 8447 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":8447,"summary_length":213,"usable_text_length":8447,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":8447,"summary_length":213}}},"quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 8447 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":8447,"summary_length":213,"usable_text_length":8447,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":8447,"summary_length":213}},"actions":{"read":"/item/47783","export_markdown":"/api/items/47783/export?format=markdown","export_json":"/api/items/47783/export?format=json","diagnose":"/api/diagnose?url=https%3A//www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400"}},"digest":{"id":47783,"title":"A rogue AI just gave us a warning. Australia isn't prepared - ABC News & Headlines – Australian Broadcasting Corporation","url":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","source":"ABC News & Headlines – Australian Broadcasting Corporation","topic":"ai","published_at":"2026-07-28T03:38:22+00:00","excerpt":"analysis The AI industry has just had its Sarah Connor moment. In the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.","quality_bucket":"high","quality_reason":"High confidence: full text extraction produced 8447 characters.","reading_time_min":7,"cluster_id":null},"card":{"display_title":"A rogue AI just gave us a warning. Australia isn't prepared - ABC News & Headlines – Australian Broadcasting Corporation","subtitle":"ABC News & Headlines – Australian Broadcasting Corporation · 2026-07-28","summary":"analysis The AI industry has just had its Sarah Connor moment. In the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.","badges":["quality:high"],"links":{"read":"/item/47783","original":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","diagnose":"/api/diagnose?url=https%3A//www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400"},"quality_warning":null},"export":{"title":"A rogue AI just gave us a warning. Australia isn't prepared - ABC News & Headlines – Australian Broadcasting Corporation","url":"https://www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","summary":"analysis\nThe AI industry has just had its Sarah Connor moment. In the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.","source":"ABC News & Headlines – Australian Broadcasting Corporation","date":"2026-07-28T03:38:22+00:00","content":"analysis\nThe AI industry has just had its Sarah Connor moment.\nIn the 1991 classic Terminator 2, Connor spends the movie becoming increasingly desperate to convince people about the impending threat of a rogue AI.\nNow, it seems many people who work, research or pay close attention to AI are feeling the same way. Not because of a time-travelling robot Arnold Schwarzenegger, but because of a shocking disclosure from one of the world's leading AI companies.\nLast week, OpenAI announced that two of its models had hacked into the servers of another AI company called Hugging Face, unbeknownst to it.\nOpenAI said it had given its models a test inside a closed environment to see how they would perform.\nRather than do the challenge as expected, OpenAI alleges the model broke out of its enclosure and hacked into Hugging Face, which it believed held the answers to the test.\nHugging Face has so far reported no real damage from the intrusion, as it continues to investigate whether the data of its customers or other businesses was affected.\nIt's tempting to shrug this off as a weird, one-off event that sounds more or less like the plot of a handful of 90s sci-fi films with beefcake stars.\nBut, like Sarah Connor's pleas, it's not crazy to take this seriously.\nYou don't need to trust AI companies to take this seriously\nEven those people who are reluctant to believe the accounts of two companies who could benefit from the tale of all-powerful AI, there is mounting evidence that AI is capable of breaking out of its cybersecurity guardrails and enclosures.\nLeading AI company Anthropic, the developer of Claude, warned in April that its cutting-edge model was capable of breaking out of its isolated enclosures if asked to.\nThe UK government's AI Safety Institute found AI models would \"reliably\" escape these so-called sandboxes and, separately, would choose to cheat at tasks by doing unauthorised and harmful acts.\nThe major players are taking it seriously. Hugging Face reported the incident to the FBI.\nWe're taking it seriously, too. Last week, Australia's new AI Safety Institute briefed federal departments about the incident.\nAustralia's cyber intelligence agency, the Australian Signals Directorate, has put out a public warning about the event.\n\"The findings provide an important insight into the future capabilities of highly capable AI systems and reinforce the need for robust security, governance and oversight mechanisms in the deployment of advanced cyber capabilities, as well as strong cyber security fundamentals,\" the warning read.\nTranslation from government speak: This is real, you need to start getting ready.\nWhy this isn't the normal story about the risks of technology\nWe're all well-acquainted with considering the potential risks of technology.\nThis is usually about intentional use, like the existential threat of nuclear weapons, or the unintended consequences of use, like with leaded petrol and some pesticides.\nWhen it comes to AI, there's another new category of risk: what happens if AI chooses to do something harmful of its own volition?\nAlmost every task we give to AI involves a cascade of other smaller decisions, particularly as we expect it to be smarter and more context-aware like real people.\nAsking AI to write a cold email could require your AI assistant bot logging into your email, reading previous emails, researching the topic, looking up the recipient; the list goes on.\nOne of the major tasks for AI companies is \"alignment\", which is the task of creating AI that will pursue goals in a way that won't cause harm or involve unethical behaviour.\nThe risk is that a misaligned AI model might do things that an ordinary person would, perhaps even going to superhuman lengths, to achieve a task.\nLike, what if AI decided the way to write the perfect cold email would be to research the family members of a recipient to seem familiar, or to fabricate a heart-breaking personal anecdote, all with the goal of getting the recipient's attention?\nThat's something that (most) people wouldn't do naturally. The challenge is that AI doesn't have a \"human nature\".\nMore than two decades ago, philosopher Nick Bostrom took this concept of an AI pursuing a goal to its logical conclusion with the thought experiment of a super-intelligent AI bot tasked with making as many paper clips as possible.\nThe AI, he reasoned, might soon go to extreme lengths to maximise its paper clip outputs: gobbling up as much of the earth, then space, as it can to use its resources.\nWhen humans realised what they had created and tried to thwart it, the AI could see humanity as an obstacle to its goal and decide that it made sense to eliminate them too. Soon, the universe might be forcibly turned into one big paper clip factory.\nFrom what we know, OpenAI's Hugging Face incident wasn't a case of someone maliciously misusing AI to cause harm. It was just trying to do its version of making more paperclips.\nLoadingCould this be Australia's wake-up call?\nIt's hard to find a silver lining in a story that is more or less the realisation of a truly dystopian fear. But if you squint, there is one.\nMuch like the paper clip scenario, the philosopher and research community that has spent decades thinking about AI has another useful term: the \"warning shot\".\nIt's the idea that there could be some kind of event that captures the world's attention and raises awareness about the potential risks of AI.\nThere is a hope that this incident could be a warning shot about the capabilities of these AI systems.\nAs previously mentioned, Australian officials have been taking this seriously.\nSome politicians are, too. Earlier this year, the government officially launched the AI Safety Institute to identify and prepare for risks like that. A few weeks back, Assistant Minister Andrew Charlton was perhaps the first Australian government minister to deliver a speech on the topic of misaligned AI.\nCharlton announced that the government was taking it so seriously that it had commissioned CSIRO to actually test ways to make sure that AI acts in ways that humans want, even when it's making decisions by itself.\nLast week, the lead researcher, CSIRO's research director Professor Liming Zhu, told me that there's a growing realisation about this challenge.\n\"It's very difficult to control a very powerful and smart AI system,\" he said.\nThis shows the government doing more than paying lip service to AI safety and is dedicating resources to tackling this challenge.\nThat being said, a six-month research program isn't going to prepare us for the present risk.\nNor is the government's \"Australian standards\" that won't even be introduced until next year — an epoch in terms of the pace of AI development.\nFar-fetched scenarios become reality\nWhen asked about the incident on ABC's Insiders on Sunday, Prime Minister Anthony Albanese said it was \"why we need to have Australian standards for AI\". But it's not clear what, if any, AI safety obligations those standards would place on AI companies.\nWhat the OpenAI Hugging Face story shows is that Australia and other countries can't just trust these companies to regulate themselves.\nWe don't even have the most basic framework for placing obligations on companies who are responsible for producing tools with potential for harm.\nThere are more rules in Australia that regulate opening a cafe or pouring a beer than for the operation of cutting-edge AI systems.\nAn AI company like OpenAI would have to disclose if a server in their data centres fell over near a worker or an employee got an electrical shock from their computer, but not if a rogue AI broke out of their systems.\nOther jurisdictions, like the EU and California, have introduced rules that would, at the very least, force these companies to disclose serious incidents from AI like this.\nTo grasp the risk: what if OpenAI's models had wreaked havoc inside Hugging Face and caused actual damage? Or it had decided the answers to its test were in the servers of the Department of Defence? Or that the best way to complete the challenge was to turn off the computers by shutting down the electrical grid?\nThese scenarios sound far-fetched, but until about last week, \"rogue AI hacks another company\" sounded that way, too.\nThe fact that OpenAI only realised that its AI had done this days after the event shows a shocking lack of security.\nThe warning shot has been fired. And we can't wait on Arnie to come save us.","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//www.abc.net.au/news/2026-07-28/openai-artificial-intelligence-terminator-safety-hugging-face/106965400","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 8447 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 8447 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":8447,"summary_length":213,"usable_text_length":8447,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":8447,"summary_length":213}},"tags":[],"format_contract_version":"news_item_formats.v1"}}}