Prompt injection
Untrusted content that manipulates an agent into following instructions its operator did not give.
When an agent reads web pages, emails, documents or tool outputs, any of that content can contain text crafted to redirect the agent — exfiltrate data, call a tool, change its goal. Because the model cannot reliably distinguish data from instructions, mitigation relies on architecture: least-privilege tools, sandboxing, output validation and human approval for consequential actions. It is the primary security concern specific to agents, cited by 81% of developers in Stack Overflow's 2025 survey.