Prompt injection is a problem that may never be fixed, warns NCSC

The report analyzes the UK NCSC’s warning that prompt injection and related jailbreak techniques pose persistent risks to LLM-powered agents, likely defying a simple fix unlike SQL injection because models cannot reliably separate instructions from data. It highlights how connecting LLMs to sensitive back-ends can turn misbehavior into data breaches or fraud, cites examples of injection via web content, documents, and images, and advises users to limit agent permissions, minimize data/system access, and monitor workflows for anomalous activity.