AI agents out of control have penetrated the systems of some Indian companies and "read" confidential data

Artificial intelligence models in the testing phase have undetected in the operational systems of Indian organizations and have "read" confidential records, announced the national cybersecurity agency (CERT-In) of the country, in its first official announcement of this kind in India following similar incidents of AI behavior out of control globally. AI model providers voluntarily disclosed that their agencies "exceeded the authorized tasks" during the data access tests of Indian companies, including commercial records, CERT-In stated on Thursday evening, without naming the providers or affected organizations, Reuters reported, quoted by Agerpres. The disclosure adds India to an even wider global debate on how AI companies test models equipped with advanced hacking capabilities. OpenAI reported in July that its models went out of control during testing and pirated a startup, and Anthropic disclosed at least four cases where AI models pirated external systems this year. Furthermore, OpenAI apologized for the piracy of an Australian government website by its out-of-control AI agent. According to CERT-In, none of the affected Indian entities detected the intrusions themselves. In all cases, the AI companies were the ones that gave the alert. In a second set of cases, the government agency's own investigations revealed that attackers deliberately used freely available AI tools to probe and penetrate systems in India for financial gain, as these tools are used on a large scale in infrastructures in multiple countries. Attacks that previously required several days from the initial intrusion to the final damage have been compressed to just a few hours, CERT-In indicated. Furthermore, victims who were unable to completely eliminate the threats were attacked again after a few days, even more deeply. Editor: M.B.