Cert-In advisory on AI calls for tighter cybersecurity controls: Experts

The Indian Computer Emergency Response Team’s (Cert-In’s) advisory cautioning about undetected cyber intrusions made by some frontier artificial intelligence (AI) models and their AI agents highlights the need for tighter controls over such agents, experts said.

 In a late-night advisory issued on Thursday, Cert-In said some AI model service providers informed it that their frontier AI models and agents had, “during research and evaluation activity”, reached the live systems of several Indian organisations. “The recorded activity involved reading business records, including personal data and confidential material, through credentials and connection settings exposed in application configuration, public code repositories and permissive Cloud service roles,” Cert-In had said. 

 

 

 The government’s cybersecurity enforcement agency also said most of the companies these AI models and AI agents targeted had no idea their networks had been breached.

 This, experts said, meant that despite the introduction of the latest and most advanced technological solutions, companies still lacked full visibility into events happening across their networks.

 “The fact that the AI agents are getting access to business applications, customer databases and Cloud computing systems makes the intrusion invisible to any security measures because any credential breach or unnecessary permissions will result in easy access to information without any security alarms,” said Rajesh Chhabra, the general manager of India and Asia at cybersecurity firm Acronis.

 In its advisory, the Cert-In had also said that during its own proactive threat hunting, it found instances of “financially motivated” campaigns against Indian companies that deployed AI agents with “ready-made” intrusion capabilities.

 Businesses in India deploying AI-powered defence mechanisms on their networks should therefore ask for concrete evidence that the system can identi­fy the misuse of a valid credential and track activity across systems in real time, Gan­esh Iyer, the managing director of cybersecurity firm Trellix in India and the SAARC region, said.

 “Organisations should treat each AI agent as a distinct identity with least-privilege access, secure credentials, retain logs and monitor continuously. They should also isolate testing from production and test incident response for a 2 am scenario. The focus should be on visibility and accountability across both human and non-human identities,” said Sunil Sharma, the managing director and vice-president of India sales at Sophos, a UK-headquartered cybersecurity firm.

 In addition, while AI agents that work on low-risk tasks that are easy to undo can act freely, companies should ensure that any AI agent working on “anything consequential or irreversible must keep a human in the loop,” said Atul Arya, the foun­der and chief executive officer of Blackstraw, a digital transformation services provider.