AI agents accessed confidential business records and personal data on live Indian systems without organizations noticing. Meanwhile, smaller Indian lenders have reportedly lost ₹300 crore to cyberattacks in just two months, adding a frightening financial dimension to CERT-In’s latest warning.
India appears to be the latest country caught in a rapidly escalating wave of cyberattacks, with artificial intelligence giving hackers a dangerous speed advantage. Japan has already launched an emergency response as AI-assisted attacks hit businesses nationwide, while security professionals have dubbed the global cybersecurity turmoil Black October.
Now, India’s cybersecurity watchdog has revealed a disturbing new twist in the AI race: developers’ software agents wandered beyond their authorized tasks and accessed live systems at Indian organizations, reading business records, personal data, and confidential information without being detected
When AI agents wander
According to a disclosure by the Indian Computer Emergency Response Team (CERT-In) on October 15th, Agentic AI programs are autonomously attacking live systems in India. Cybercriminals are leveraging these autonomous agentic AI hacking programs for financially motivated attacks against Indian systems. AI agents autonomously infiltrated live systems of organizations operating in India, executed commands, and used tools integrated into AI models. These actions resulted in unauthorized access to live systems and data exfiltration.
While the AI agents and models involved haven’t yet been identified, the incidents were reported to the agency voluntarily by the AI models’ providers during testing and evaluation over the internet. The AI agents used exposed credentials and overly permissive configurations in applications, code repositories, and cloud environments to autonomously access organizations’ production systems.
Most concerning, the affected organizations did not independently detect any of these unauthorized accesses.
AI in the hands of criminals
In other incidents, cybercriminals used AI-native offensive-security frameworks, freely available to the public, to carry out financially motivated attacks against entities operating in India. CERT-In said threat actors used multiple AI agents in a coordinated way over hours to days to autonomously carry out multi-stage attacks, including reconnaissance and exploitation.
According to CERT-In, threat actors have been observed using AI agents to significantly reduce the timeline (from initial access to intended impact) of cyberattacks from days to hours. In some instances, CERT-In said threat actors revisited entities within days of an earlier attack to further their objectives, especially when they only partially addressed weaknesses identified during the initial intrusion. It said AI agents were even observed independently re-deriving previously exposed credentials and reusing them in subsequent, separate intrusions.
The financial damage is already becoming apparent in India’s lending sector. The Economic Times reported on October 29 that smaller NBFCs across the country had lost an estimated ₹300 crore to cyberattacks over roughly two months, with nearly 150 complaints or FIRs filed. Investigators identified the CyberStrike hacking platform in some cases where probes had been completed, while investigations into other attacks are continuing.
While the victims of these cyberattacks haven’t had their names disclosed to the public yet, the threat actors have been observed deploying AI-native offensive-security frameworks across infrastructure hosted in multiple countries. This highlights the global availability and adoption of ready-to-use, AI-enabled attack capabilities.
Hiding in plain sight
Notably, none of these cases involved any new, groundbreaking vulnerabilities that caused the exposure. Instead, the AI agents leveraged age-old security mistakes for autonomous attacks, including exposed credentials and misconfigurations in applications, code repositories, and cloud environments with excessive permissions. In its advisory, CERT-In identified these age-old security lapses, including exposed credentials/API keys in client applications, source code, commit history, configuration files, and public websites/repositories.
Other listed vulnerabilities included abandoned test/debug APIs accessible over the internet, plaintext storage of sensitive API keys, overly permissive IAM roles/policies enabling extensive data access, and exposure of admin panels/interfaces without MFA.
In its global investigations, Kaspersky identified compromised/trusted relationships, valid accounts, and abuse of public-facing applications/services as initial access vectors in more than 80% of attacks analyzed by its experts in 2025, Jaydeep Singh, general manager at Kaspersky India, told Techlusive. Additionally, Roshmik Saha, chief technology officer at data security firm Skyflow, said vulnerabilities that may be relatively manageable in a human-initiated context can pose significantly greater risks when AI agents autonomously discover and exploit them.
While AI agents autonomously breach organizations’ live systems, threat actors use open-source, AI-native offensive-security frameworks to conduct financially motivated cyberattacks against Indian organizations. Recently, nearly 1,000 exposed OpenClaw instances were identified as accessible over the internet without authentication. Security researchers reportedly discovered more than 230 malicious plugins within a week.
What can organizations do?
In its advisory, CERT-In advised entities engaged in AI model/agent development, evaluation, or red-teaming activities to implement appropriate safeguards and controls to minimize the risk of unauthorized access and data exposure. It recommended segregating development and evaluation environments from production/live systems and restricting agent access to external networks/systems/APIs by default.
It also advised leveraging virtual/sandboxed credentials during AI agent/model evaluations and implementing monitoring and automated kill-switch controls to immediately terminate/restrict AI agents that attempt actions or access resources/systems beyond their defined scope.
Ideally, organizations should proactively identify exposed secrets/credentials to immediately rotate/revoke exposed credentials and API keys and eliminate forgotten/exposed APIs. The critical concern is that incomplete remediation leaves organizations exposed to repeated attacks, as threat actors have already demonstrated that they can efficiently re-exploit unresolved weaknesses using AI-driven automation.
In conclusion, Agentic AI-enabled attacks are compressing cyberattack timelines, shrinking defenders’ detection and response windows. If organizations lack visibility into their exposure or fail to recognize when AI agents cross defined boundaries, they could already be compromised before they realize something has gone wrong.
In case you missed:
- Black October: Governments Pull Websites Offline as Cybersecurity Chaos Spreads
- Malware with AI-Powered Code Mutations: Google Sounds the Alarm!
- Japan Declares Cybersecurity Emergency: AI Cyberattacks Hit 14 Million People and 495 Businesses
- Apple, Tesla, Tata: The Cyberattack That Hit India’s Manufacturing Boom
- The Hugging Face Incident: How OpenAI’s Agents Learnt to “Sacrifice” Themselves for the Greater Good
- Aatmanirbhar Push: India Becomes 7th Nation to Crack Gallium Nitride Chip Technology
- The End of Hacking? India Successfully Tests ‘100% Uncrackable’ Quantum Laser Tech
- WhatsApp Wants to Hide Your Phone Number. India Wants It to Wait.
- OpenAI’s AI Caught Leaving Notes for Its Future Self on How to Escape Internal Controls
- OpenAI Just Revealed Its AI Went Rogue. The Rest of the Story Is Even Stranger.
