Agentic AI in cybersecurity: what financial organisations should control
Agentic AI can plan tasks and call tools with limited supervision. In security operations, permissions, logging and human approval determine how safely it is used.
Agentic AI systems can plan multi-step tasks and use tools or APIs. In cybersecurity, that can help defenders automate triage, but it increases the importance of least privilege, sandboxing, audit logs and human approval.
What makes an AI system agentic
A conventional chatbot mainly responds to prompts. An agentic system can plan a sequence, call tools, retrieve data or trigger workflows. The term describes a capability pattern, not a guarantee that a system is fully autonomous or reliable.
Why financial firms need controls
A tool-using system may access customer records, identity services, payment workflows or security consoles. Excessive permissions can turn a mistaken instruction or compromised agent into a larger incident. CERT-In's November warning makes the distinction between defensive automation and attacker automation relevant.
A practical checklist
Give each agent the minimum permissions required, isolate tool execution, log prompts and tool calls, rotate secrets, require human approval for payments or destructive changes, monitor outbound traffic and test failure paths. Do not place production credentials in prompts.
Frequently asked questions
Does agentic AI mean no human involvement?
No. It means the system can plan and use tools; permissions and approval gates still control its actions.
What is the first control to implement?
Least-privilege access, isolated tool execution and complete audit logs.