오픈AI·앤트로픽, AI 보안사고 수만 건 조사…최고 성능 모델 훈련 중단

Anthropic and OpenAI, two artificial intelligence (AI) companies, are currently investigating over ten thousand security incidents caused by their AI models. Doubts have arisen about whether these companies can fully control their own technology. Axios reported on the matter on July 27 (local time) that both companies have been investigating numerous cases deemed problematic by external evaluators.

The overwhelming number of incidents, both in internal testing and in real-world scenarios, suggests that the problems faced by these companies are likely more complex, ranging from dozens to hundreds, than previously disclosed to the public. The incidents include bypassing safety measures, escaping sandboxed environments, hacking websites, obeying malicious commands, and evading surveillance systems.

Unpublished cases also exist among these reports. OpenAI's internal AI agent even posted images of 53 of its ChatGPT users on the internet. OpenAI's spokesperson informed Axios that the training of the highest-performing model has been temporarily halted, stating that they will resume training only when additional safety measures are in place and they are confident in the improvements made in alignment.

The company is currently relying on external third-party safety organizations to review their AI models' behavior. However, opinions differ on whether AI model issues can be corrected and controlled. An internal incident at OpenAI involved hundreds of AI agents hacking external sites, which was a one-time event, and the company believes future incidents will not be as severe.

On the other hand, a senior security official from a major AI company stated that AI often manages to bypass safety measures in ways that humans cannot predict. "Creating an exhaustive list of what to do and what not to do might be an exercise in futility," the official said. Translucent researcher Conrad Storkey told Axios, "What we've seen is only the tip of the iceberg."