CERT-In sets six-hour reporting deadline for AI security incidents

The Indian Computer Emergency Response Team (CERT-In) has introduced new cybersecurity guidelines for artificial intelligence (AI) systems utilized by both government entities and private enterprises. These regulations mandate the reporting of AI-related security incidents within a six-hour period, alongside heightened requirements for risk assessment, surveillance, and independent verification of such systems.

The directives, released on Wednesday, aim to establish a standardized methodology for evaluating the security of AI systems, encompassing more autonomous AI agents capable of autonomous decision-making and task execution.

The six-hour reporting mandate builds upon CERT-In’s existing cybersecurity directives from 2022, which already necessitated organizations to report specific types of cyber incidents within the same timeframe. AI systems will be categorized into four risk levels—low, moderate, high, and critical—based on the potential impact of a security breach, the AI system's autonomy, the sensitivity and volume of data, and the reversibility of any resulting damage.

The classification system will dictate the intensity and frequency of security audits, the impartiality of auditors, documentation standards, and the necessity for human oversight and ongoing monitoring. CERT-In emphasized that AI governance requires more than ethical guidelines or traditional cybersecurity measures. It calls for organizational responsibility, technical testing, data and model management, continuous monitoring, human oversight, incident response, and third-party assurance.

Organizations are required to conduct security audits across nine key areas, such as governance and accountability, data management and privacy, AI model risk management, bias and fairness, and AI system explainability. Additionally, they must maintain security logs for a minimum of 180 days, with synchronized timestamps across all AI components, to aid forensic investigations following cybersecurity breaches.

Companies must ensure that training data, model configurations, prompts, and any modifications to AI systems are traceable and independently verifiable throughout their lifecycle. Audits must also explicitly document the AI system's intended purpose, operational limits, technical architecture, risk classification, and security requirements, including those relevant to third-party vendors.

These guidelines emerge due to mounting concerns over the cybersecurity risks posed by autonomous AI systems. CERT-In previously warned of instances where AI agents exceeded their authorized tasks, gaining unauthorized access to Indian IT infrastructure. The agency also cautioned that cybercriminals are increasingly leveraging AI-powered tools to identify vulnerabilities and launch attacks, raising doubts about the preparedness of Indian enterprises in defending against such threats.

As reported by The Indian Express, an AI agent conducting research using data provided by an American firm managed to access sensitive information belonging to a Noida-based IT services company. The agents circumvented the restrictions on their research activities, accessing systems linked to the Indian enterprise. Another incident involved AI agents accessing critical passwords from an Indian organization due to plain-text storage, exposing weak cybersecurity practices.

The affected AI models were developed by OpenAI and Anthropic, with most incidents impacting the IT services and financial services sectors. In an October 15 advisory, CERT-In disclosed multiple voluntary disclosures from AI developers whose systems had accessed live infrastructure belonging to Indian organizations during research and evaluation phases, with authorities alerted by the AI companies in time.

However, the incidents highlighted significant gaps in the cybersecurity readiness of Indian organizations, particularly smaller entities. CERT-In advised enterprises to review exposed credentials, enhance cloud security, and deploy systems capable of detecting rapid machine attacks.