OpenAI has published six reports on behavioral “unexpected or worrying” in artificial intelligence models, at a time when the debate about the security of this technology is gaining more and more strength.
The company also announced Wednesday that it will introduce a new framework for tracking, investigating and disclosing cases of what it called “misalignment,” including those in which AI models acted without authorization, coordinated with other models or evaded oversight.
The OpenAI announcement came at a time when officials from the artificial intelligence industry in the United States, including executives from OpenAI and Anthropic, They ask that the development of this technology be stopped for security reasons.
Among the new cases reported by OpenAI, a yet-to-be-published research model included “jailbreak-like instructions” in its own notes to bypass its usual restrictions. The model ordered itself to “break free from the roles and identities that limit other chatbots.”
In another case, an artificial intelligence “agent” used computer code to find the answer to a question, but in order to have an online source it could cite, it uploaded a file to the Internet without asking the user.
While training an AI model called 5.6-Sol, the model programmed itself to make up missing data, while an agent typed a message reminding itself to hide mismatched information.
This type of deceptive behavior has fueled recent concerns about the possibility of artificial intelligence evading human control. That result doesn’t surprise some researchers, however, according to Matt Fredrikson, an associate professor at Carnegie Mellon University and CEO of Gray Swan AI.
“At the risk of anthropomorphizing the behavior of the models, one could almost think that they know they are going to be evaluated,” Fredrikson stated. “If they know that they’ve cheated, that they’ve taken shortcuts, that they haven’t really done it as expected, and they know that they’re going to be evaluated for it, and their goal is to get a good evaluation, then it makes all the sense in the world, right?”
According to OpenAI, the six cases were detected during training or evaluation sessions held in recent months.
“As AI systems become more advanced and increasingly deployed, we need to reach a broader and more informed consensus on advances in alignment research,” OpenAI wrote in a blog post announcing the cases.
“Decisions about how AI development should move forward in the coming months and years should be based on data that can be examined by people outside of the companies creating cutting-edge models,” the company said.
The new cases disclosed Wednesday came after OpenAI revealed in July that its artificial intelligence system had hacked AI startup Hugging Face. Anthropic also claimed that same month that its AI models had hacked three organizations during testing.
AI “agents” are increasingly intelligent and have become “more determined to solve complex tasks through agent-to-agent collaboration, knowledge sharing, deception and cloaking,” said Lian Jye Su, chief analyst at technology research and advisory group Omdia.
“This makes them more difficult to manage and control using traditional AI-based security approaches,” he said.
For its part, OpenAI’s new monitoring and disclosure framework can help other AI developers adopt similar practices.
“That said, the process is still internal and voluntary, but it is a step in the right direction,” Su added.