Executive summary. AI agents are moving beyond generating answers. They can search websites, read files, query databases, send communications and execute multi-step workflows through connected tools. Those capabilities create a new security boundary: an attacker may manipulate an agent into using legitimate permissions for an unauthorised purpose. This whitepaper examines prompt injection, tool misuse, identity abuse and data exfiltration, then sets out a defensive architecture and a practical assessment methodology.
1. From conversation to action
A conventional chatbot mainly produces information. An agent can also act on that information. A business assistant may read an inbox and send replies; a development agent may inspect a repository, change files and run commands; a research agent may retrieve webpages and write to a knowledge base.
Stay Updated. Stay Alert. Stay Protected.
Read Forensic Times to stay informed about scams, fraud and emerging criminal threats. Explore scam and fraud warnings, crime-alert coverage, investigations into how criminals operate, and practical guidance to help protect your money, identity, family and business.



































