AI agents hit 9 in 10 Indian firms in past 1 year, only 36% dealt with it in 24hrs
An AI agent is software that can work toward a goal by taking steps, rather than only producing a response to a prompt. Depending on its design, it may use connected tools or systems to complete a task. This is a general definition; the report does not define the term in detail. A chatbot that only answers questions typically returns text to the user. An agent might instead be set up to retrieve information from a company system or perform an approved task through a connected tool. The key difference is its ability to act using those connections, not simply the label given to the software. That ability makes oversight important. ServiceNow’s survey found that 92 per cent of Indian organisations surveyed had experienced a security incident linked to AI agents in the previous 12 months. The report calls for agents to be registered, assigned an owner and made visible, so security teams can understand and monitor what they do.
What is an AI agent, and how is it different from a chatbot that only answers questions?
An AI agent is software that can work toward a goal by taking steps, rather than only producing a response to a prompt. Depending on its design, it may use connected tools or systems to complete a task. This is a general definition; the report does not define the term in detail.
A chatbot that only answers questions typically returns text to the user. An agent might instead be set up to retrieve information from a company system or perform an approved task through a connected tool. The key difference is its ability to act using those connections, not simply the label given to the software.
That ability makes oversight important. ServiceNow’s survey found that 92 per cent of Indian organisations surveyed had experienced a security incident linked to AI agents in the previous 12 months. The report calls for agents to be registered, assigned an owner and made visible, so security teams can understand and monitor what they do.
How many surveyed organisations had an AI-agent-related security incident, and how many contained one within 24 hours?
The survey by ServiceNow found that 92 per cent of the 500 Indian security leaders’ organisations surveyed experienced a security incident linked to AI agents in the last 12 months. The result shows how common these incidents were among the organisations covered by the study.
Rapid containment was much less common. Only 36 per cent of organisations were able to contain such an incident within 24 hours, according to the report as covered by PTI. Containment means stopping or limiting an incident; the report’s headline figure does not say what type of incident occurred or describe the response used in each case.
The figures sit alongside wider gaps in oversight. For example, just 19 per cent had one named person accountable for agent security, and only 18 per cent had unified visibility of their agents through a single interface. The report points to a need for stronger governance and monitoring, but does not give a forecast of future incident rates.
What gaps in ownership, visibility, and pre-deployment checks did the report find?
The report found gaps in three parts of agent security: ownership, visibility and checks before deployment. Clear ownership matters because teams need to know who is responsible for governing an agent. Yet only 19 per cent of surveyed organisations had one named person accountable for agentic AI security.
Visibility was also limited. Just 18 per cent could see their AI agents through one unified interface. For pre-deployment checks, only 30 per cent had an automated gate that stopped an agent from going live until it passed security checks. That gate is a mechanism for making checks a condition of deployment, rather than relying only on later monitoring.
The evidence process had gaps too: 47 per cent could provide a complete evidence package on demand. The report also found that half of respondents rated their agentic security capabilities as established, despite the shortcomings it identified. Together, the figures suggest that confidence and basic governance arrangements did not always match. The report argues for ownership, visibility and automated controls to keep pace with deployment.
What can go wrong when an AI agent has access to company systems but is not being monitored?
An AI agent connected to company systems can have access beyond a conversation window. As a general cybersecurity point, what it can reach or do depends on the permissions and connections it has been given. If its activity is not monitored, unusual behaviour may be harder for a security team to spot quickly.
For example, an agent might have access to a business system as part of its assigned task. If it then behaves unusually, monitoring and response controls can help a team detect and contain the issue. The report found that 51 per cent of organisations had automated, real-time detection of unusual agent behaviour, but 67 per cent could not respond automatically when an agent behaved anomalously.
The report does not specify the data exposed or actions taken in particular incidents. It does say that 92 per cent of surveyed organisations experienced an AI-agent-linked security incident in the last 12 months, while only 36 per cent contained one within 24 hours. It stresses continuous oversight and swift responses.
What is an unregistered or “shadow” AI agent, and how might one end up running in a company?
An unregistered, or shadow, AI agent is an agent operating in a company environment without being recorded or known to the security team. The report found that 37 per cent of surveyed organisations had discovered AI agents running in production during the past 12 months that had never been registered or known to their security teams.
The report does not say how those agents got there. In general, an organisation might end up with an untracked tool if a team deploys software without completing a central registration process, but that is a possible route, not a cause identified by this study. The report’s finding is about agents already running in production, not just proposed tools.
An unknown agent can leave security teams without a clear inventory, owner or view of its activity. The report found that only 19 per cent had one named person accountable for agent security and only 18 per cent had unified visibility. It recommends registering each agent, assigning an owner and making it visible in one place.
How do an AI agent’s permissions determine what company data it can reach and what actions it can take?
Permissions are the access rules that determine which information or functions a software tool can use. As a general cybersecurity principle, an agent connected to company systems can reach only the resources its permissions and integrations allow. Those rules therefore shape what company data it can access and what actions it can take.
For example, an agent given access to a particular business system may be able to use information or functions available through that connection. An organisation can limit the potential impact of misuse by restricting access to what is needed for the task. The article does not detail specific permission settings or give an example of an agent accessing data improperly.
The report stresses that access alone is not enough to understand risk: knowing what an agent can reach is not the same as seeing what it is doing. It found that only 18 per cent of organisations had unified visibility of their agents. Registration, ownership, automated controls and continuous monitoring can help teams govern both access and activity.
What basic cybersecurity controls help organisations spot unusual activity, limit damage, and respond to an incident?
Organisations need controls at different stages: before an agent goes live, while it is operating and when something unusual happens. The report specifically highlights registering each agent, assigning an owner, making agents visible in one place, using automated checks before deployment and maintaining continuous oversight.
A deployment gate is one preventive control: it stops an agent from going live until it clears security checks. Monitoring can then flag unusual behaviour in real time, while response controls help teams contain an incident. In the survey, 30 per cent had an automated deployment gate and 51 per cent had automated, real-time detection of unusual activity.
Response was a weak point. Although 51 per cent could detect unusual behaviour automatically, 67 per cent could not respond automatically when an agent behaved anomalously. Only 36 per cent contained an agent-linked incident within 24 hours. The report calls for automated controls and swift responses, alongside clear ownership and visibility, to help security teams spot and contain issues.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web