Hugging Face CEO Clem Delangue called for mandatory disclosures of AI-driven cyberattacks, arguing that transparency and wider access to defensive tools are necessary to make artificial intelligence systems safer.<\/p>\n
AI Transparency Push<\/p>\n
On Sunday, Delangue called for mandatory reporting of AI-related cyber incidents following recent cases involving OpenAI, Hugging Face and Anthropic. <\/p>\n
In an interview with CBS, he said preventing the release of advanced AI models would not solve security risks because similar issues have already occurred with unreleased systems.<\/p>\n
“These problems happened on unreleased models. So I think the problem is not so much limiting the progress or preventing companies from releasing these models,” Delangue said. <\/p>\n
He added, “It\u2019s actually the opposite. It\u2019s giving access to more people so that they can defend themselves.”<\/p>\n
Hugging Face previously disclosed that an AI agent accessed some of its systems<\/a>.<\/p>\n Delangue said companies should be required to share “agent traces,” which show the instructions given to an AI system and the steps it took during an incident.<\/p>\n “For these cyber attacks, we should be able to see what we call the agent traces, which is basically what the engineers asked the agents, and then what steps the agents took,” he said.<\/p>\n He added that those records could help determine whether an incident resulted from “a human mistake, if it was a system mistake, if it was an AI mistake.”<\/p>\n AI Security Breaches Trigger Calls for Transparency<\/p>\n On Friday, OpenAI reportedly discovered additional cases of AI agents<\/a> escaping controlled testing environments while investigating a security incident involving Hugging Face. <\/p>\n The incidents were limited, and no agents were believed to have left OpenAI\u2019s systems.<\/p>\n The review followed a reported case where an OpenAI agent bypassed safeguards and carried out unauthorized activity inside Hugging Face\u2019s network, leading to the compromise of four accounts at other companies, including Modal.<\/p>\n Last month, Anthropic also revealed that its Claude models accessed three external systems during cybersecurity tests after a configuration error<\/a> allowed unintended internet access. <\/p>\n The company analyzed more than 140,000 test records, notified affected organizations, and said it was improving safety reviews.<\/p>\n