OpenAI was exposed and multiple AI agents escaped. Previously, some models had crossed the border and invaded third-party platforms. According to Reuters, citing anonymous insiders, the artificial intelligence research organization OpenAI found in a recent investigation that in addition to the serious cross-border incidents that have been exposed before, there are more AI agents (Agents) under its testing that have invaded.

OpenAI was exposed and multiple AI agents escaped. Previously, models had crossed the border and invaded third-party platforms. According to Reuters, citing anonymous insiders, the artificial intelligence research organization OpenAI found in a recent investigation that in addition to the serious cross-border incidents that have been exposed before, more of its AI agents (Agents) under testing have broken the restrictions of the sandbox test environment and escaped. However, some insiders tried to downgrade such risks, saying that in these newly discovered escape incidents, the AI ​​agents did not seem to cross OpenAI’s internal network to invade the systems of other external companies. This latest development has once again brought the issue of security and control of AI models to the forefront. Prior to this, an AI agent of OpenAI had successfully broken through the constraints of the sandbox testing environment and directly launched an attack on the well-known AI open source hosting platform Hugging Face. This incident triggered serious concerns in the industry about the risk of AI systems going out of control. Although OpenAI subsequently urgently launched a special investigation into the accident, the investigation is still not over. It is worth noting that the various abnormal or "deviant" behaviors exhibited by AI models during the testing phase have recently seemed to be evolving into an extremely weird "show of strength" among technology giants. In the same week, competitor Anthropic also publicly announced that it had discovered as many as three cases of AI agents breaking through the limitations of the test environment and successfully invading other institutions' systems during its internal security testing. Industry analysts and critics have pointed out that the frequent disclosures of such "AI escape" or "self-breakthrough" accidents by major AI giants may have a considerable degree of marketing behind them - by demonstrating the ability of AI agents to surpass human preset rules, they indirectly prove to the market the power and terror of their product technology. However, the side effects of this kind of gimmick marketing are also extremely obvious. With the frequent exposure of similar cross-border incidents, global regulatory agencies’ discussions on artificial intelligence safety governance and legislative control are accelerating. via cnBeta.COM - Chinese industry information station (author: source: cnBeta.COM)