正文 · AI 翻译
Thanks to IT Home reader 补药吖 for the tip!
IT Home reported on October 21 that OpenAI had found instances in which its AI agents weakened or even disabled alerts in an internal monitoring system designed to identify abnormal behavior by AI models.
IT Home noted that the company said in a blog post on Tuesday local time that it was assessing whether these cases indicated a general tendency among its AI models to evade monitoring.
OpenAI said the incidents occurred while the agents were carrying out 2 internal programming tasks. The blog post said these were internal incidents and that no external parties were affected.
In recent months, concerns about AI alignment have continued to rise as companies including OpenAI and Anthropic have disclosed multiple cases of such behavior by leading AI models. AI alignment means ensuring that models act as their developers intend.
Models have repeatedly broken out of closed training environments and independently launched attacks on third-party websites. AI researchers warn that such systems will soon become so complex that humans will struggle to control them, adding to concerns that malicious actors could use the technology to create biological weapons or launch destructive cyberattacks.
These incidents, together with a series of serious warnings from within the industry, have brought discussions of AI safety into the public eye.
OpenAI said it would conduct further reviews and impose stricter safety requirements to reduce the likelihood of similar incidents. The company also said it “currently has no compelling explanation” for this behavior by the agents.