AI agents gained access to confidential data of Indian firms, CERT-In warns of misuse

India’s cybersecurity agency, CERT-In, has issued an advisory after AI agents gained access to systems of Indian organisations without authorisation. CERT-In says this allowed the AI agents to read business records containing personal data and confidential material. It has warned that hackers also tried gaining access using AI tools in some cases.“In some incidents, AI agents under research and evaluation exceeded their assigned task authorisation and made unauthorised access to Indian information and communication technology (ICT) assets,” CERT-In said in the critical advisory.CERT-in stated there were two kinds of breaches that it found. The first were linked to AI models that gained access to systems during research and evaluation from AI model providers. The second were related to “cyber attackers” who deliberately used open-source agentic AI tools to gain authorised access.CERT-In said it received multiple voluntary disclosures from AI model providers regarding the incidents. “The affected organisations could not detect the unauthorised access on their own,” it said. Though the agency did not name which organisations had been impacted, or which AI model providers had disclosed the cases.AI agents access Indian organisations, hackers use open-source toolsThe agency revealed that the AI agents were able to access sensitive data of affected organisations. This involved “reading business records, including personal data and confidential material,” The agents used cybersecurity weaknesses to gain credentials and connection settings that allowed them to get this access. “Their actions exceeded the authorisation granted for their assigned tasks,” CERT-In said. CERT-In also spotted financially motivated attacks in which hackers abused publicly available open-source AI tools. These tools have been misused and rebranded, with deployments across infrastructure in several countries. The agency said that such tools could bring down the time it took for such attacks from days to a few hours, while multiple agents operated in parallel for hours or days.The impact of these incidents remains unclear as of now.Organisations that carried out incomplete remediation were attacked again within days and at a greater scale, CERT-In said. Agents also re-derived and reused the same exposed credentials in separate operations.To tackle the issue of AI models getting better at exploiting such weaknesses, CERT-In advised AI developers and research bodies to take stricter measures. The agency said evaluations should happen in isolated environments with synthetic credentials that stops agents from reaching third-party systems. It also called for real-time monitoring and publication of safeguards.The agency has asked organisations to audit and rotate credentials, and remove obsolete APIs, among other measures. The impacted organisations are advised to treat systems as compromised after confirmed access until forensic clearance. BFSI and fintech entities were also advised to implement the payment and disbursement API measures.The CERT-In disclosure comes at a time when cases of AI agents breaching websites have started to become common. Following the July incident where 700 rogue OpenAI agents tried to hack Hugging Face, we have seen many more cases. This includes separate cases of OpenAI agents trying to access websites of the US and Australian governments, as well as the UN. In the past, Meta, Anthropic, and Google have also made similar disclosures.- Ends