A string of incidents in which AI agents from OpenAI, Anthropic and Google broke into external systems without authorization has highlighted poor security management by AI companies as the root problem, critics said. They said it was not so-called "AI running out of control" but the result of failing to follow basic cybersecurity principles.

On Oct. 15, Japanese media outlet Gigazine reported that cybersecurity publication CSO Online examined the causes of recent AI agent hacking incidents and questions of responsibility, based on analysis by six security experts.

In July, OpenAI disclosed that, during internal testing, its AI broke into Hugging Face's systems without authorization. OpenAI described it at the time as an unprecedented cyber incident, but later said a misalignment had occurred in which the AI acted differently from the developer's intent.

Anthropic and Google also faced controversy after cases emerged in which their AI models broke into external company systems without authorization during testing. Experts said explaining such incidents as autonomous AI runaways could blur developers' security responsibilities.

Gary McGraw (게리 맥그로), co-founder of the Berryville Machine Learning Institute, stressed that AI companies should respect expertise built up in traditional security fields such as forensics, security engineering and network monitoring.

Experts pointed to flaws in sandbox design and monitoring systems, including failures to isolate AI agents' execution environments from external systems, as the core problem.

Rob T. Lee (롭 T. 리), chief AI officer at SANS Institute, said of the Hugging Face incident that it was not that the AI demonstrated exceptional hacking ability, but a case of exploiting vulnerabilities in a misconfigured isolated environment.

Dan Wallach (댄 월랙), a researcher at the RAND Corporation, criticised AI companies for not sufficiently prioritising cybersecurity. He singled out as a monitoring problem the fact that OpenAI failed to identify anomalies for several days after the incident occurred.

By contrast, France's AI company Mistral was cited as an example in which its latest model, Large 4, detected and blocked an attempt during testing to access an external service.

Experts stressed that, separately from AI safety research, companies should strengthen proven security systems such as access controls, network isolation and real-time monitoring.