The Indian Computer Emergency Response Team (CERT-In) flagged AI asymmetry as a key risk to the BFSI and payments ecosystem in its Digital Threat Report 2025-26, warning that offensive capabilities of frontier AI models are scaling faster than regulatory, defensive and operational frameworks designed to contain them.
India’s BFSI sector suffered 2.9 million cyber attacks in 2025, increasing over 2x from 1.4 million in 2021.
The report, copy of which is with MediaNama, points to two specific instances over the last 12 months to demonstrate how AI tools have automated significant parts of the cyber threat landscape.
- November 2025: Anthropic disclosed GTG-1002, where a Chinese-linked cybercriminal group used its Claude AI tool to target 30 companies globally. AI carried out up to 90% of the operation, in what was described as the first large-scale cyberespionage campaign largely run by an AI system.
- April 2026: Claude Mythos Preview autonomously discovered over 23,000 vulnerabilities across more than 1,000 open-source projects, out of which 1,752 were high- or critical-severity vulnerabilities.
CERT-In released its first advisory on AI threats in 2023, flagging that AI models could generate malicious code and phishing content.
Cyberattacks are becoming systemic: AI moves adversaries from manual, specialist operations to repeatable, automated attack pipelines, the report says.
- Vulnerability discovery, exploit chaining, and payload delivery are getting industrialised more like software manufacturing than tradecraft.
- GTG-1002 demonstrated this directly: the agent scaled, firing thousands of requests per second across roughly 30 organisations, including financial institutions.
AI is changing cyberattack economics: AI makes targeted campaigns cheaper to run, increasing the likelihood of a probe into any given organisation.
- Ransomware groups, which previously relied on buying exploits (master keys) from brokers to break into the systems of their targets, can now generate them independently using AI, at a fraction of cost.
- Frontier models have demonstrated autonomous vulnerability discovery across operating systems, browsers, and, directly relevant to digital asset operations, smart contracts, where frontier models produced working attacks against 207 of 405 historical smart contract exploits, totaling $550 million in simulated stolen funds.
Three types of AI-driven attacks that are already happening in BFSI
1. Parallel, multi vector campaigns: Hackers are using AI to launch massive, simultaneous attacks against a bank’s internal network, employee emails, cloud systems, and third-party vendors.
“Incident response built on sequential attack assumptions is structurally unprepared,” the Digital Threat Report 2025-26 said.
2. Discovery-to-exploit timeline is shrinking: The window between a hacker discovering a vulnerability and actually exploiting is compressing sharply. AI allows attackers to find flaws at “machine speed.”
While the attack velocity is rising, financial institutions remain constrained due to internal red tape—change in management, compliance gates, and approval processes.
3. Hijacking the software pipeline: Cybercriminals are increasingly using AI to attack DevOps practices that automates the software development lifecycle rather than attacking the app.
What BFSI institutions must do now
1. Set up real-time inventory tracking for old legacy systems, unmanaged endpoints (personal devices employees use for work), accounts without multifactor authentication and remote access tools that lack proper security.
2. Stop hiding passwords in code: Developers sometimes accidentally leave passwords, API keys, or sensitive credentials inside the software code or in the automated systems that build the software (CI/CD variables). Hackers look for these. Completely remove these “secrets” from the code.
3. Make security checks automatic, not an annual event: Usually, banks check their software for flaws (Static and Software Composition Analysis) once a year or during a big audit. These checks must be built directly into the software development pipeline so that code is scanned for flaws automatically every single time a developer makes a change.
4. Patch vulnerabilities immediately: Waiting weeks or months to install software updates leaves the bank open to attacks. Use “virtual patching”, tools that automatically block hackers from exploiting a flaw the second it is discovered.
5. Detection beyond signatures: AI-generated exploits frequently
lack prior patterns in threat feeds. Financial institutions should test security systems to ensure they can spot anomalous behaviour, even if the specific attack method has never been recorded before.
6. Adopt the ‘zero-trust’ operating model: In the GTG-1002 campaign, the AI agent did not need to hack the network, as it already held valid keys required to operate from within. Continuously verify every user, device, and identity, both human and non-human, across your systems.
This comes less than a month after CERT-In released a blueprint in May, outlining how organisations should defend against AI-assisted cyber threats.
Also read:
