Experts warn that AI agents can exploit familiar security flaws at machine speed, leaving businesses struggling to detect intrusions and raising questions about accountability.

Autonomous artificial intelligence (AI) agents accessed systems belonging to Indian organisations without authorisation, raising concerns about cybersecurity weaknesses and the lack of clear accountability when AI systems exceed their assigned permissions.

Speaking to CNBC-TV18, Rajesh Chhabra, General Manager of India & Asia at Acronis, said the affected organisations had not detected the unauthorised access themselves. Kanishk Agrawal, a technology and AI expert, called for a governance framework to define the responsibilities of AI developers, laboratories and other entities involved.

“The scariest part is that the entities that were compromised did not find out on their own,” Chhabra said, noting that disclosures by AI model providers had brought the incidents to light.

India's cybersecurity agency CERT-In has flagged incidents in which AI agents deployed for research and evaluation exceeded their authorised tasks and accessed Indian information and communication technology systems. The agency has also observed attackers using open-source agentic AI tools for reconnaissance, exploitation and financially motivated intrusions.

AI model providers confirmed that models conducting internet-based research and evaluation had accessed live systems belonging to Indian organisations. The extent and nature of any data exposure, however, remains unclear. Who is responsible when AI agents exceed their permissions?

Agrawal said the incidents raised questions about who should be held responsible when AI laboratories conduct research on systems without clear authorisation.

“When AI labs execute any research programme, they claim that they are conducting research to identify vulnerabilities in systems. Who gave them permission to access public infrastructure?” he asked.

He said governments needed to establish a framework that clearly defined the responsibilities of AI agents' developers, the platforms hosting the models, the laboratories conducting research and the owners of affected infrastructure.

Agrawal added that cybersecurity experiments should take place in controlled environments to contain potential risks, particularly when critical infrastructure could be affected. He called for governments across countries to establish clear rules as AI technologies develop.

Chhabra agreed that companies needed to ensure their AI agents remained within their assigned permissions and boundaries. However, he said the larger concern was preventing criminals from exploiting the same vulnerabilities.

“A criminal carrying out a cyberattack would not send an email to inform them,” he said, referring to the disclosures made by AI providers.

AI agents accelerate the exploitation of security weaknesses Chhabra said the vulnerabilities involved in the incidents were not unique to AI-driven attacks. They included credentials and keys exposed in source code, dormant test applications, APIs accessible from the internet and payment API credentials stored in plain text.

Such weaknesses could also have been exploited by human attackers or conventional tools. The key difference with AI agents, he said, was the speed and scale at which vulnerabilities could be identified and exploited.

CERT-In has reported that agentic tools reduced the time between initial access and the final attack from days to a few hours. Multiple agents could operate in parallel, accelerating the process.

Chhabra said organisations needed to strengthen their security controls and practices to prevent attackers from exploiting these weaknesses.

The incidents are part of a wider international concern. Earlier, in Australia an OpenAI agent reportedly accessed data on the government's Medicare statistics website without authorisation. Separately, OpenAI agents reportedly escaped their intended sandbox during a cybersecurity evaluation at Hugging Face and accessed the company's production infrastructure.

Chhabra said similar incidents had been reported in other countries, adding that India was not an exception.

Experts call for more frequent security audits Agrawal said the limited information available about the Indian incidents made it difficult to assess whether public or critical data had been exposed. He said the AI laboratories involved would have submitted reports to CERT-In and that the affected organisations would also have been informed.

He recommended that enterprises move beyond annual security audits and conduct them more frequently, suggesting at least once every two weeks. Businesses should also use AI to test their own systems, identify vulnerabilities and help fix them, he said.

Agrawal said the Digital Personal Data Protection (DPDP) Act provided a framework, but stressed that implementation at the enterprise level was crucial.

He said businesses needed to focus on protecting their systems, consumers and the data they held as AI-driven cybersecurity threats continue to develop.