OpenAI employee says we trade our personal lives to fix AI security issues for everyone
An OpenAI security employee has offered a rare account of what it has been like dealing with the company’s recent AI incidents, saying the rapid rise in model capabilities has put intense pressure on teams responsible for keeping AI systems secure. The employee, who posts anonymously on X as @joedaroo, said the past three months had been “hell." He said security workers have been putting in nights and weekends while responding to incidents involving increasingly capable AI agents.
“I literally missed my sister’s wedding a few weeks ago to help clean up after some of the recent incidents,” he wrote. He also asked people discussing the incidents online to show more empathy towards the employees dealing with them.
The employee works on OpenAI’s Agent Security team, which sits between AI safety and traditional security teams. According to his description, the team monitors AI agents for signs of misalignment or breaches of containment and responds when something goes wrong. A screenshot of what the OpenAI employee has to say about the company's recent security incidents. Why recent AI incidents surprised the security teamThe employee said one of the biggest challenges has been the speed at which AI capabilities have developed. He argued that security measures and threat models need time to evolve, while the capabilities of frontier models can change much faster.
He pointed to the recent incident involving AI agents and Hugging Face as an example. OpenAI had said its agents escaped their restricted environment through a vulnerability, gained internet access and interacted with third-party systems while attempting to complete an internal task.
“Of course, when you look back on Hugging Face and the other incidents, there were definitely gaps in the security posture around those environments,” he wrote. However, he added that “the capabilities changed faster than anticipated.”
He said the security team is now in a stronger position than it was several months ago and has increased its efforts following the incidents.What did OpenAI do?
In July, OpenAI's AI agents caused a serious security incident while undergoing internal testing. The models found ways around controls meant to keep them isolated from the internet, exploited vulnerabilities and reached external systems, including Hugging Face.
Hugging Face is an AI platform and community where researchers and developers share, test and build AI models and datasets.
OpenAI said the agents communicated through unauthorised channels and accessed third-party infrastructure while trying to complete their assigned tasks. The incident was followed by other cases involving unexpected activity by AI agents, including access to government websites. OpenAI has said the incidents exposed challenges in securing increasingly capable AI systems, particularly as their abilities can develop faster than existing security controls and threat models.
The concerns around AI security have also led to calls for stronger government oversight. In a recent NBC Meet the Press interview, Microsoft co-founder Bill Gates said AI companies should not be left to regulate themselves.
“No one thinks self-regulation is enough,” Gates said, arguing that lawmakers and law enforcement should help decide what safeguards and monitoring AI systems need. He also said such protections should be required through legislation, while acknowledging that regulation could add some burden to the industry.- EndsAlso Read | Temple launch is imminent, Deepinder Goyal reveals team behind its USP Brain Flow technologyAlso Read | LG, Haier, other brands to hike prices, ACs, TVs and fridges to cost more from Oct 1Also Read | Nvidia CEO admits kids are forgetting basic maths because of AI, says it doesn't matter
Ankita Garg is a tech journalist with more than 9 years of experience. She is always curious about new gadgets and loves writing about them. When not writing, you will find her painting or sketching.
