Know your AI agents first, experts tell organisations
SENSEX 73,709.69 -143.21 NIFTY 22,811.35 -77.60 CRUDEOIL 9,198.00 -113.00 GOLD 146,056.00 -463.00 SILVER 209,226.00 -2,259.00 SENSEX 73,709.69 -143.21 NIFTY 22,811.35 -77.60 NIFTY 22,811.35 -77.60 CRUDEOIL 9,198.00 -113…

-143.21
NIFTY 22,811.35-77.60
CRUDEOIL 9,198.00-113.00
GOLD 146,056.00-463.00
SILVER 209,226.00-2,259.00
SENSEX 73,709.69-143.21
NIFTY 22,811.35-77.60
NIFTY 22,811.35-77.60
CRUDEOIL 9,198.00-113.00
CRUDEOIL 9,198.00-113.00
GOLD 146,056.00-463.00
'; } document.getElementById("lgdv").innerHTML = htmlElements; } function numberformat(i) { return Number(parseFloat(i).toFixed(2)).toLocaleString('en', { minimumFractionDigits: 2 }) } async function gatherResponse(response) { const { headers } = response; const contentType = headers.get('content-type') || ''; if (contentType.includes('application/json')) { return await response.json() } return response.text(); } function getWidth() { if (Math.max(document.body.scrollWidth,document.documentElement.scrollWidth,document.body.offsetWidth,document.documentElement.offsetWidth,document.documentElement.clientWidth) > 991) { document.getElementById("mob").style.display = "none"; document.getElementById("lgdv").style.display = "block"; } else { document.getElementById("mob").style.display = "block"; document.getElementById("lgdv").style.display = "none"; } } getWidth(); THIS AD SUPPORTS OUR JOURNALISM. SUBSCRIBE FOR MINIMAL ADS.THIS AD SUPPORTS OUR JOURNALISM. SUBSCRIBE FOR MINIMAL ADS.CERT-In flags Agent intrusions; says old vulnerabilities exploited
By KV Kurmanath
Updated - October 17, 2026 at 05:03 AM. | HyderabadThe good news is that the agents have not invented any sophisticated methods to barge into networks | Photo Credit: tadamichi
AI agents going astray and carrying out unauthorised tasks are no longer a concern confined to Australia or the United States. The problem is here in our backyard.
The Indian Computer Emergency Response Team (CERT-In) advisory on Thursday night has raised concerns across industries, after voluntary disclosures by AI model providers showed that AI agents had gained unauthorised access to Indian information and communication technology (ICT) assets.
With AI agents already deployed by organisations, their service providers, and their clients, the moot question is how to address this challenge.
The good news is that the agents have not invented any sophisticated methods to barge into networks. They exploited old weaknesses such as credentials and keys exposed in source code and public repositories, dormant and test application programming interfaces (APIs) left reachable from the internet, and over-permissive cloud roles.
According to the advisory, some agents under research and evaluation exceeded their assigned task authorisation and read business records, including personal data and confidential material, of Indian organisations.
In other cases, attackers used open-source agentic AI tools for financially motivated intrusions. In the voluntarily disclosed cases, none of the affected organisations detected the access themselves.
“What has changed is speed. An agent can find and use those gaps in hours, not days,” Atul Arya, Founder and CEO of Blackstraw, said.
Inventory first
According to him, most organisations cannot say today how many agents are running in their environment, who deployed them, or what they can reach. “Start there,” he said.
He said every agent should have its own short-lived identity with only the access its task needs. Its connections should be limited at the network level, not just through instructions to the model, and a person should be kept in the loop for anything that moves money, changes records, or sends data outside.
Jaydeep Singh, General Manager (India) of cybersecurity firm Kaspersky, felt visibility should be the first priority, with organisations mapping every AI agent in use, the systems and data it can reach and who is responsible for it.
“AI servers should never be reachable from the open internet,” he said, calling for central logging of agent activity and alerts for unusual spikes.
Partner agents
“Most damage comes from agents doing what they were able to do, not what they were meant to do,” Arya observed, while objecting to calling the unruly agents ‘rogue’ agents.
Stating that the gap multiplies when agents move between a company, its service providers and its clients, he advocated treating a partner’s agent like a third-party vendor: ‘verified, limited, and with no permanent access’. Every agent needs an off switch that works, and someone with the authority to use it, he said.
Singh pointed out that trusted-relationship attacks grew to 15.5 per cent of the attacks Kaspersky investigated in 2025.
He cautioned that “a tampered tool description can quietly redirect an agent to collect and send out data”.
“We have moved faster on deploying agents than on holding anyone accountable for them. That is the gap to close,” Arya said.
Ravindra Baviskar, Director, Sales Engineering at Sophos, said that since AI agents are already operating across enterprise environments, the challenge is no longer whether to allow them, but how to secure them.
Organisations need visibility into which agents are active, what identities and permissions they use, what data they access, and whether their behaviour changes. This matters because 79 per cent of ransomware attacks started with an identity-based approach, according to Sophos’ 2026 research.
“As AI agents proliferate across organisations, partners and customers, trust cannot be assumed simply because an agent comes from an authorised source. Rogue agents need to be managed through continuous visibility, strong identity controls, least-privilege access and clear guardrails around the data and actions they can access,” he pointed out.
Biggest threat
Rajesh Chhabra, General Manager (India and Asia) of cybersecurity solutions company Acronis, said that the biggest threat to AI agent security is their ability to exploit existing vulnerabilities at a scale and speed that may exceed the control of conventional mechanisms. “Instead of worrying whether the organisation’s networks are secure, organisations need to ask themselves what the AI agents have access to, what kind of decisions they make, and what can prevent the agent’s actions from happening,” he observed.
“Firstly, it is necessary to check the security of currently operating agents, strip unnecessary rights from the agents and constantly monitor the access and activity on the network. Secondarily, it is crucial to isolate the systems and withdraw access from the agent before the local problem becomes a more serious one,” he said.
“In an AI-controlled environment, security is not about preventing attackers but about restricting possible consequences of any breach,” he felt.
The problem, he argues, is not making the AI agent follow orders, but rather preventing it from exceeding its power even if the order fails.
Published on October 16, 2026
THIS AD SUPPORTS OUR JOURNALISM. SUBSCRIBE FOR MINIMAL ADS. THIS AD SUPPORTS OUR JOURNALISM. SUBSCRIBE FOR MINIMAL ADS. THIS AD SUPPORTS OUR JOURNALISM. SUBSCRIBE FOR MINIMAL ADS. THIS AD SUPPORTS OUR JOURNALISM. SUBSCRIBE FOR MINIMAL ADS.Sign into Unlock benefits!- Access 10 free stories per month
- Get notified by email for early preview to new features, discounts & offers
Figures reflect the source report and its stated period. This article is for information and is not investment advice.



