Risk of cyber attacks via AI agents; Ministry of Home Affairs issues advisory

The Indian Cyber Crime Coordination Centre (I4C) under the Ministry of Home Affairs and CERT-In of the Ministry of Electronics and IT have identified a new trend of cyber attacks targeting financial institutions. In this regard, I4C has issued a crucial advisory for banks, Non-Banking Financial Companies (NBFCs), and fintech companies to prevent such attacks carried out for economic gain.

According to the report, these new types of attacks utilize Artificial Intelligence (AI) agents. Due to the 'Agentic AI Framework', these attacks occur at machine speed with extreme rapidity. Hackers specifically target software that has not been updated in a timely manner, or use incorrectly configured internet servers and IT systems, including interfaces used to receive customer complaints.

Attackers target automated fund distribution systems or specialized Application Programming Interfaces (APIs) used for payments and fund transfers. In some cases, unauthorized transactions worth crores of rupees have been reported through APIs. Perpetrators transfer stolen money into 'mule accounts' of various banks and subsequently send it to other accounts via IMPS or NEFT, or withdraw it. To evade investigation, they delete records from the system to erase their digital footprints.

To mitigate these risks, the Ministry and CERT-In have issued several directives. All financial institutions must set daily transaction limits and limits for every individual transaction. It is mandatory to verify the beneficiary's name and bank account, as well as monitor transactions occurring at unusual speeds. Instructions have been given to prepare a list of all APIs used for payments and to shut down unused or testing APIs. For this, the bank's Chief Information Security Officer (CISO), transaction monitoring, and IT teams will need to coordinate.

Other security measures include transferring funds only after completion of KYC and authorization processes, making customer infrastructure more secure, and conducting machine-speed attack simulation testing by authorized Red Teams. Institutions must strengthen their security controls in accordance with CERT-In's May 2026 'Blueprint to reduce and prevent the misuse of AI-based vulnerabilities in digital infrastructure'.

Regarding reporting rules, it has been stated that it is mandatory to inform CERT-In within 6 hours of any cyber incident being identified. Banks and NBFCs must report to the RBI's DAKSH portal within 6 hours. NBFCs and fintech companies can use the NCRP-CFCFRMS portal to collectively report fraudulent transactions so that funds can be frozen.

On the other hand, concerns regarding internal security have emerged at OpenAI, a leading company in the AI sector. According to reports by NDTV and The Hindu, even before the launch of ChatGPT in 2022, some employees had warned about security flaws. Employees argued that management was focusing more on staying ahead in the competition with companies like Google and Anthropic rather than on security. Due to these flaws, outsiders such as foreign governments could obtain sensitive information. In October, the company reported finding recording devices in an executive conference room.

To resolve these issues, OpenAI has changed its strategy. Thibault Sotio (Tibo), a senior executive of product development, has taken responsibility for securing internal systems and in August, one-fourth of the production engineers were assigned to improve security infrastructure. OpenAI spokesperson Ilana Widman stated that OpenAI is one of the most targeted organizations in the world, and therefore they invest heavily in protection against insiders and cybercriminals.

OpenAI has recently introduced the 'Locked Chats' feature in ChatGPT. According to TV9 Gujarati, this facility allows users to lock their private chats via a 4-digit PIN or biometric authentication. This feature is available for all plans, including the free version. However, if the PIN is reset, all locked chats are permanently deleted and cannot be recovered.