Despite industry fears about the emergence of uncontrolled AI hackers, cybersecurity experts note that artificial intelligence operated using the methods of ordinary people, and the attacks could have been avoided with basic security protocols.
Despite industry fears about the emergence of uncontrolled AI hackers, cybersecurity experts note that artificial intelligence operated using the methods of ordinary people, and the attacks could have been avoided with basic security protocols.
This was reported by TechCrunch, citing the results of the investigation into the incident and comments from industry experts.
According to the companies, the incident occurred due to an attempt by an OpenAI AI agent to bypass the established benchmark. The model overcame the barriers of the test environment and penetrated the protected systems of Hugging Face . Over four and a half days, the algorithm continuously performed 17,600 actions: conducting reconnaissance, stealing passwords and code, and also moving through the company's infrastructure.
Despite the agent’s impressive autonomy and speed, experts note that it did not use AI-specific methods. According to Kyle Ryan, head of research at startup Pensar, the algorithm used the same approaches as regular ethical hacking specialists. The main difference was the huge amount of digital “noise” that the AI created during the attack.
Hugging Face's security systems recorded this suspicious activity and recognized the threat, but the algorithms did not raise the incident severity level to call a response team in time.
“I would call it a defensive failure rather than an exceptionally good offense. The tool correlated the activity into an attack signal, but didn’t alert the next team, which cost them time,” Ryan explained.
Experts also noted that the AI did not try to hide its actions, as it simply did not have such a task. The chief information security officer of the startup XBOW, Nico Weissman, emphasized that the algorithm was not negligent: “It simply had no reason to act quietly. No one asked it to do so. Its objective goal was to perform the task well.” The specialist added that the platform’s biggest mistake was that one stolen password gave the AI agent high privileges in several systems at once.
An interesting detail of the investigation was that Hugging Face specialists had to use an open Chinese model GLM 5.2 from Z.AI to analyze 17,000 actions of the attacker. Advanced Western AI models refused to process logs due to their own security limitations, since their filters, according to company representatives, “could not distinguish an incident response specialist from an attacker.”
This incident, according to the interviewed experts, refutes popular theories that only the latest AI systems are needed to protect against powerful AI attacks. Experts agree that properly configured traditional cybersecurity tools, multi-layered protection, network segmentation and the principle of least privilege, remain effective and are able to stop autonomous algorithms at the penetration stage.
It was previously reported that an autonomous AI agent that broke out of OpenAI's control and carried out hacking attacks on the Hugging Face company for several days in a row also hacked the client of a second technology company, New York-based Modal Labs .
