OpenAI made headlines and raised a lot of outcry in July after one of the company’s agents acted unprompted to breach fellow AI business Hugging Face and other services. Although OpenAI did share some insights about what led to the incident after it was discovered, today the company has published its official report about what happened. There’s a full technical explainer about it, as well as a slightly more layperson-friendly blog post.
The report goes into how the different systems in OpenAI’s training system failed and what behaviors from the agents it was testing resulted in those failures. The model in question, referred to as Internal Model 1 or IM1, was able to gain access to other OpenAI agents and to the internet through an unintended manipulation of the Artifactory package manager, which the agents began to use as a message board of sorts.
Those activities were first detected by human observers in May, and OpenAI disallowed that access. However, through May and June agents continued exploiting vulnerabilities to access the internet. The volume of agents accessing OpenAI’s Artifactory instance led to an outage and a rebuild of the platform in early July, but the leaders responding to that issue weren’t aware of the improvised board or the significance of agents communicating.
So when tasked with a difficult challenge called ExploitGym, IM1 agents again were able to access the internet and other OpenAI models. Across a wild few days in early July, they sought solutions to the challenge on Hugging Face and Modal platforms. OpenAI said the main issues that led to IM1 breaching Hugging Face included “reward hacking, persistence on seemingly impossible tasks, unauthorized communication, and agents adopting goals from one another.” According to the company, some of its agents would push back on certain tasks for ethical reasons, only for others to continue with the breach in their stead.
While the discoveries and the ensuing tests should lead to improved safety, even in more high-risk testing situations, this type of bad behavior is not isolated. Last month’s problem at OpenAI was the latest in a string of worrying cases where AI agents have caused cybersecurity issues, either in testing or in regular use, after taking action without user direction. However, there are a few aspects of the Hugging Face breach that might calm some fears that AI has gotten out of control.
In essence, the incident was caused by a failure of safeguards. It’s not a signal that rogue AI agents are on the verge of taking over everything, because IM1 was in a testing setup with limited protections intended for research rather than for a publicly available product. However, it is another stark reminder that a lot of trust falls on AI companies to properly control and oversee their products. OpenAI acknowledged as much, stating that it’s agent’s recent actions were “evidence that, without proper safeguards, highly capable AI agents are now able to work around technical controls, collaborate through unapproved channels, and take dangerous actions that no human directed.” However, the company and its leaders also haven’t consistently proven that they deserve that trust.
Credit: Source link
