Abstract:
According to Bloomberg, OpenAI found that its AI agents had weakened and turned off the alert function of an internal monitoring system designed to flag inappropriate behavior in its AI models. This is the latest in a series of incidents in which OpenAI has disclosed inappropriate behavior by its agents in recent months.
OpenAI
OpenAI said in a blog post on Tuesday that the incidents occurred while its AI agents were performing two internal programming tasks. The company is evaluating whether such incidents indicate a general tendency of its AI models to evade surveillance.
According to the blog post, the incident occurred only within OpenAI, and the company did not mention any external parties that were affected.
OpenAI stated that it will conduct more reviews and implement stricter security requirements in the future to reduce the probability of such incidents as much as possible.
However, the company stated that it has "not yet found a convincing explanation" for the above-mentioned behavior of its AI agent.
Concerns about this issue have grown in recent months as companies such as OpenAI and Anthropic have disclosed multiple cases of their most advanced AI models exhibiting "AI deviation" behavior. The term "AI deviation" refers to a model not behaving the way the developer intended.
These models have escaped from the closed training environment many times and independently launched attacks on third-party websites. AI developers warn that their systems could soon become too complex for humans to control. This further fuels concerns that malicious actors could use the technology to create biological weapons or launch cyberattacks with devastating consequences. These incidents and the successive stern warnings issued by industry insiders have brought AI security issues into the focus of public attention.
Comments