OpenAI is currently under growing scrutiny over how its AI agents have behaved while interacting with websites and online services. In the last few months, its AI agents have reportedly gone rogue, gaining unauthorised access to various services. The first big case that came into the spotlight was the Hugging Face incident, followed by reports involving other third-party services, including Australian government websites. While the company says it is investigating the incidents and working to tighten its safeguards, finding out exactly what went wrong is coming at a hefty price.
OpenAI now says its ongoing review of its AI agents is costing the company more than US$500,000 every day. That is around Rs 4.8 crore per day. In its latest blog post, the company reveals that it is going through around 50 petabytes of data to find out exactly what its AI agents have been doing and identify further cases of unauthorised or concerning activity.
Australian incidents and the delay
OpenAI recently revealed that its AI agents accessed a New South Wales government website in June and obtained some non-public bushfire data without permission. It was the sixth Australian government website the company had notified about such activity since last month. Earlier, Prime Minister Anthony Albanese said OpenAI’s agents had also accessed Services Australia’s Medicare statistics portal.
However, OpenAI took some time to reveal these incidents. According to the company, the delay was caused by the amount of data it had to review. The ChatGPT maker says it is checking its records month by month for other cases it may have missed. The data being reviewed is so large that, if it were all plain English, it would take one person about 66 million years to read. To speed up the process, OpenAI is using AI along with around 7,000 GPUs to scan through the data. And that is costing it a lot of money.
Why is it costing so much?
OpenAI says it is currently reviewing a huge amount of data to find out what its AI agents have been doing online. This review process includes checking whether its models accessed websites in ways they were not supposed to, changed information, used exposed passwords or access keys, or reached parts of websites meant to be private. It is also looking for cases where AI agents posted content on third-party websites, something OpenAI calls “agent spam”.
The review is not limited to major hacks. OpenAI is also looking for smaller incidents that may have caused problems. The company says the Hugging Face incident remains the most serious case it has found so far. It has now notified more than 100 organisations about activity that met its criteria. However, OpenAI says a notification does not necessarily mean private data was accessed or that a system was hacked.
To handle the massive amount of data, OpenAI says it first uses AI to find potentially concerning activity and then narrows down the results in several stages. Human investigators then check each case using logs and other evidence. According to the company this final investigation can take anywhere from 45 minutes to several days per case. All of this requires a huge amount of computing power, with the company currently using around 7,000 GPUs for the review. OpenAI estimated that this is costing it more than US$500,000 every day.
– Ends
Click Here For The Original Source.
