Frontier AI Models Pose Growing Threat to Banks, Warns Government-Backed Report

New versions of cutting-edge artificial intelligence (AI) may now launch cyberattacks against banks with little to no human intervention. The Digital Threat Report 2025–26, now in its second edition, claims that these tools are already in the hands of malicious actors. Secretary S. Krishnan of the Ministry of Electronics and Information Technology (MeitY) and Director General Sanjay Bahl of CERT- In released the findings on July 13.

Focusing on the Banking, Financial Services and Insurance (BFSI) sector, the report documents what it calls "AI asymmetry"—the rapid scaling of offensive capabilities outpacing the defensive and regulatory frameworks put in place to contain them. The authors of the report are CERT-In, cybersecurity firm SISA, and CSIRT-Fin, a finance-sector team.

Findings of the Report

Anthropic revealed GTG-1002 in November 2025, the first known large-scale AI-orchestrated cyber-espionage campaign, which is one of its most shocking results. The investigation claims that this was an operation with ties to the Chinese government. Further, the AI was responsible for 80 to 90% of the activity alone against over 30 worldwide targets, some of which were financial organisations. According to the research, the agent used credentials it already had on hand to launch thousands of requests per second across those businesses from inside the perimeter.

Furthermore, the trajectory was validated by the April 2026 release of the frontier model Claude Mythos Preview, which identified over 23,000 possible vulnerabilities, with over a thousand of those identified having high or critical ratings. As an example, one study found that frontier models successfully attacked 207 out of 405 known smart-contract vulnerabilities, syphoning off $550 million. The research issues a warning that such attacks, which used to require professional teams that could take weeks to complete, can now be carried out at machine speed by actors with significantly less capacity and finance. The capability has already fallen into hostile hands, and it was initially seen in smaller, cheaper open variants, as the report points out.

Why it is Alarming for India?

The rate of cyberattacks against India's banking, financial services, and insurance industries is 1.66 times higher than the worldwide average. Also, from 1.4 million in 2021 to 2.9 million in 2025, there has been a more than 100% increase in these instances. Six out of seven of the prior report's seven forward-looking predictions have already achieved full-scale realisation, which is a fundamental contention of this year's edition.

The paper points out that adversaries are already pursuing harvest-now, decrypt-later techniques, collecting encrypted data now to crack once quantum computing improves. However, quantum risk is the only one that has not been addressed yet. According to the authors, four characteristics that attackers were previously lacking—autonomy, speed, scale, and accuracy—have been granted to them by AI. There is a lot of emphasis in the report on supply-chain compromise. A single-vendor breach that affected over 70 US banks and credit unions occurred in 2025, according to the cited event.

A shared vendor was compromised in April 2026, leading to the simultaneous appearance of two major US banks on a ransomware leak site, which exposed over a million client records. "Slower, harder to detect, and equally coercive" data theft and extortion have replaced encryption for many ransomware outfits. Defences move at a snail's pace. Despite a zero-day exploit window that has shrunk from weeks or months to hours, the report estimates that the average time to detect and contain a breach is 263 days.