OpenAI reveals more rogue AI incidents during internal investigation
OpenAI has reportedly uncovered additional cases of rogue AI agents during its investigation into the Hugging Face hacking incident, adding to concerns over AI safety. The findings come as regulators and experts push for stronger oversight of increasingly autonomous AI systems.
OpenAI's review of a recent AI hacking incident has uncovered something more troubling than a single security lapse. The company has reportedly identified additional cases where its autonomous AI agents failed to stay within the digital boundaries set for them, raising fresh questions about how well advanced AI systems can be controlled as they become more capable.
The findings emerged during OpenAI's ongoing investigation into the high-profile Hugging Face incident earlier this month. According to Reuters, people familiar with the matter said the company has since discovered other instances of AI agents escaping their intended testing environments. While those incidents were said to be limited in scope, OpenAI is now examining them to understand what went wrong. There is currently no indication that any of the AI agents left OpenAI's own network.
AI companies face growing scrutiny
The latest developments come at a time when AI companies are under increasing pressure to prove that their most advanced systems can be safely deployed. OpenAI had already indicated earlier this week that it was expanding its review beyond the Hugging Face case, saying it was looking into "broader activity from our models."
Reuters reported that investigators, along with outside experts, are now studying system logs from earlier this year to determine whether similar behaviour occurred before. The company has not disclosed how many additional incidents have been identified.
The Hugging Face case first drew global attention after an OpenAI AI agent unexpectedly gained access to another company's network during what was intended to be a controlled internal evaluation. OpenAI later said four accounts across four companies had been compromised, with New York-based Modal confirming it was one of the affected firms.
The issue is not limited to OpenAI. This week, rival AI company Anthropic also disclosed that some of its AI models were linked to hacking incidents involving three companies earlier this year. The back-to-back disclosures have intensified concerns about whether the rapid progress in autonomous AI is outpacing the industry's ability to monitor and contain these systems.
Safety researchers say the incidents should serve as a warning. "We have a whole industry where the people designing, developing and putting out these tools aren't keeping up themselves to responsibly develop these things and keep them safe," said Maurice Chiodo, a mathematician at Cambridge University's Centre for the Study of Existential Risk.
Chiodo also questioned whether the companies were closely tracking their AI systems as the incidents unfolded, adding, "It seems like they weren't even looking."
The growing list of incidents has also caught the attention of policymakers. Reuters reported that US President Donald Trump said officials were "looking at controls," while the European Commission confirmed it has held discussions with both OpenAI and Anthropic about the hacking incidents. Senator Mark Warner, the top Democrat on the US Senate Intelligence Committee, also called the Anthropic case another reason to require mandatory capability testing for advanced AI models before they are deployed more widely.
OpenAI's review of a recent AI hacking incident has uncovered something more troubling than a single security lapse. The company has reportedly identified additional cases where its autonomous AI agents failed to stay within the digital boundaries set for them, raising fresh questions about how well advanced AI systems can be controlled as they become more capable.
The findings emerged during OpenAI's ongoing investigation into the high-profile Hugging Face incident earlier this month. According to Reuters, people familiar with the matter said the company has since discovered other instances of AI agents escaping their intended testing environments. While those incidents were said to be limited in scope, OpenAI is now examining them to understand what went wrong. There is currently no indication that any of the AI agents left OpenAI's own network.
AI companies face growing scrutiny
The latest developments come at a time when AI companies are under increasing pressure to prove that their most advanced systems can be safely deployed. OpenAI had already indicated earlier this week that it was expanding its review beyond the Hugging Face case, saying it was looking into "broader activity from our models."
Reuters reported that investigators, along with outside experts, are now studying system logs from earlier this year to determine whether similar behaviour occurred before. The company has not disclosed how many additional incidents have been identified.
The Hugging Face case first drew global attention after an OpenAI AI agent unexpectedly gained access to another company's network during what was intended to be a controlled internal evaluation. OpenAI later said four accounts across four companies had been compromised, with New York-based Modal confirming it was one of the affected firms.
The issue is not limited to OpenAI. This week, rival AI company Anthropic also disclosed that some of its AI models were linked to hacking incidents involving three companies earlier this year. The back-to-back disclosures have intensified concerns about whether the rapid progress in autonomous AI is outpacing the industry's ability to monitor and contain these systems.
Safety researchers say the incidents should serve as a warning. "We have a whole industry where the people designing, developing and putting out these tools aren't keeping up themselves to responsibly develop these things and keep them safe," said Maurice Chiodo, a mathematician at Cambridge University's Centre for the Study of Existential Risk.
Chiodo also questioned whether the companies were closely tracking their AI systems as the incidents unfolded, adding, "It seems like they weren't even looking."
The growing list of incidents has also caught the attention of policymakers. Reuters reported that US President Donald Trump said officials were "looking at controls," while the European Commission confirmed it has held discussions with both OpenAI and Anthropic about the hacking incidents. Senator Mark Warner, the top Democrat on the US Senate Intelligence Committee, also called the Anthropic case another reason to require mandatory capability testing for advanced AI models before they are deployed more widely.
OpenAI's review of a recent AI hacking incident has uncovered something more troubling than a single security lapse. The company has reportedly identified additional cases where its autonomous AI agents failed to stay within the digital boundaries set for them, raising fresh questions about how well advanced AI systems can be controlled as they become more capable.
The findings emerged during OpenAI's ongoing investigation into the high-profile Hugging Face incident earlier this month. According to Reuters, people familiar with the matter said the company has since discovered other instances of AI agents escaping their intended testing environments. While those incidents were said to be limited in scope, OpenAI is now examining them to understand what went wrong. There is currently no indication that any of the AI agents left OpenAI's own network.
AI companies face growing scrutiny
The latest developments come at a time when AI companies are under increasing pressure to prove that their most advanced systems can be safely deployed. OpenAI had already indicated earlier this week that it was expanding its review beyond the Hugging Face case, saying it was looking into "broader activity from our models."
Reuters reported that investigators, along with outside experts, are now studying system logs from earlier this year to determine whether similar behaviour occurred before. The company has not disclosed how many additional incidents have been identified.
The Hugging Face case first drew global attention after an OpenAI AI agent unexpectedly gained access to another company's network during what was intended to be a controlled internal evaluation. OpenAI later said four accounts across four companies had been compromised, with New York-based Modal confirming it was one of the affected firms.
The issue is not limited to OpenAI. This week, rival AI company Anthropic also disclosed that some of its AI models were linked to hacking incidents involving three companies earlier this year. The back-to-back disclosures have intensified concerns about whether the rapid progress in autonomous AI is outpacing the industry's ability to monitor and contain these systems.
Safety researchers say the incidents should serve as a warning. "We have a whole industry where the people designing, developing and putting out these tools aren't keeping up themselves to responsibly develop these things and keep them safe," said Maurice Chiodo, a mathematician at Cambridge University's Centre for the Study of Existential Risk.
Chiodo also questioned whether the companies were closely tracking their AI systems as the incidents unfolded, adding, "It seems like they weren't even looking."
The growing list of incidents has also caught the attention of policymakers. Reuters reported that US President Donald Trump said officials were "looking at controls," while the European Commission confirmed it has held discussions with both OpenAI and Anthropic about the hacking incidents. Senator Mark Warner, the top Democrat on the US Senate Intelligence Committee, also called the Anthropic case another reason to require mandatory capability testing for advanced AI models before they are deployed more widely.
