What AI Agent Attacks Mean for India’s Cyberspace

On October 15, India disclosed its first cyber incident where an AI agent, not a human operator, was responsible. The Government of India issued new guidance on November 4 for auditing AI systems and reporting future incidents. With frontier AI development concentrated in the United States and AI diffusion running through India and the Global South, New Delhi and Washington need to coordinate more in this area. As Prime Minister Modi and President Trump prepare to meet twice before year-end, there is a window for both countries to develop a blueprint for reporting, governing, and containing future incidents.

Q1: What happened?

A1: In its October 15 “critical” severity advisory, India’s Computer Emergency Response Team (CERT-In) identified two distinct types of AI agent incidents against Indian technology assets. The first involved AI agents developed by frontier labs that exceeded the scope of their assigned evaluation tasks during model research and development, escaped containment, and accessed live Indian systems, similar to the mid-2026 OpenAI–HuggingFace incident. It is a serious but tractable threat, but is addressable through stronger sandbox design and authorization controls. The second type of attack is more consequential: criminal actors deliberately weaponizing open-source agentic AI tools to conduct “reconnaissance, exploitation, and financially motivated intrusions,” compressing attack timelines from days to hours.

Q2: Why does this matter?

A2: In every incident CERT-In documented, the breach came to light either through voluntary disclosures from AI providers or through CERT-In’s own threat hunting. None of the affected organizations detected the intrusions, even as agents accessed and read personal data and confidential business records.

Legacy network infrastructures, fragmented security controls, limited automation, and a talent gap mean only 7 percent of Indian organizations have achieved a “mature” cybersecurity readiness rating. AI is compounding these gaps: 68 percent of Indian organizations have limited or no AI and automation in security operations, and those without it take an average of 311 days to identify and contain a breach, 55 more than those with extensive AI deployment.

India is racing toward a $1 trillion digital economy, and sees AI as a key accelerant. Sustaining that growth will require investor confidence in India’s ability to govern AI risks as adeptly and quickly as it adopts AI.

Q3: How has the Indian government responded?

A3: The Indian government has been proactively engaged on the threat of AI-enabled cyber. CERT-In’s Comprehensive Cyber Security Audit Policy Guidelines, published in July 2025, established a structured audit framework for critical infrastructure; in April 2026, CERT-In issued a high-severity advisory (CIAD-2026-0020) flagging autonomous reconnaissance and multistage attack orchestration as systemic threats. Finance Minister Nirmala Sitharaman convened the country’s private and public banking executives to level-set on advancing cybersecurity standards after news of the cyber capabilities of Anthropic’s Mythos class of models became public in May 2026. In July 2026, the Reserve Bank of India issued a directive asking banks and nonbanking financial entities to submit a board-approved assessment of their cyber threat landscape. The October 15 CERT-In disclosure has now accelerated these efforts.

CERT-In’s November 4 AI Security Auditing Guidelines represent India’s most comprehensive response yet. They call for all AI systems, including those operated by foreign developers, to be classified into four risk tiers: low, moderate, high, and critical. Low-risk systems can self-assess, while moderate- and high-risk systems face annual independent audits. The guidelines also outline requirements for high- and critical-risk systems to face a combination of emergency kill switches, continuous monitoring, and red-team exercises, depending on their tier rankings. For agentic AI systems, they call for all systems to remain within “approved operating parameters” to address sandbox failures. CERT-In also extends its mandate to report all cybersecurity incidents within six hours of any AI-related incidents and demands that enterprises maintain system logs for 180 days.

Q4: How can the United States and India work together to address AI-enabled cyber vulnerabilities?

A4: The labs building the most capable AI systems are overwhelmingly American, but as Indian users and enterprises engage with these systems as early adopters at scale, they will be the ones increasingly absorbing their risks. That gap creates a shared obligation in the U.S.-India relationship. The Quad Leaders’ Summit in Manila and the G20 Leaders’ Summit in Miami in December give President Trump and Prime Minister Modi two near-term opportunities to advance cooperation on reporting, governing, and containing AI-enabled cyber incidents moving forward.

  • Reporting: When Presidents Trump and Xi met in September, they agreed to a bilateral U.S.-China AI incident communication channel. If Washington is building that with Beijing, it should build a more substantive equivalent, with defined thresholds and operating protocols, with allies and partners, including New Delhi.
  • Governing: The U.S. Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) is developing AI agent standards and conducting predeployment cybersecurity evaluations with frontier labs. CERT-In’s November 4 guidelines show India’s desire to address audit domains and risk classification, but do not directly address these areas. Establishing mechanisms to align U.S. standards with India’s audit framework should be a near-term deliverable for U.S. and Indian policymakers.
  • Containing: Both governments bring a myriad of agencies to responding to agentic AI-enabled cyber incidents: for the United States, the National Security Council, CAISSI, and the Cybersecurity and Infrastructure Security Agency; for India, the National Security Council Secretariat, the nascent but not yet fully operational AI Safety Institute, CERT-In, and the Indian Cybercrime Coordination Centre. With technology as new as agentic AI, the rules are being written in real time, with limited shared visibility into how those agencies interact within each country and across borders. Mapping that architecture, developing shared standard operating procedures for cross-border incidents with designated roles for each agency, and scheduling joint tabletop exercises and wargaming scenarios between these agencies should be a near-term deliverable that the National Security Advisor–led Transforming the Relationship Utilizing Strategic Technology initiative should seek to deliver.

Aman Thakker is a fellow in the Chair on India and Emerging Asia Economics at the Center for Strategic and International Studies (CSIS) in Washington, D.C. Aditya Goswamy a program coordinator and research assistant in the Chair on India and Emerging Asia Economics at CSIS.

Image
Aman Thakker
Fellow, Chair on India and Emerging Asia Economics
Image
Aditya Goswamy
Program Coordinator and Research Assistant, Chair on India and Emerging Asia Economics