An attack that hides instructions in text an AI system reads, tricking it into ignoring its original rules or taking unintended actions.
Language models treat instructions and data as the same kind of input: text. Prompt injection exploits this. An attacker places instructions in a web page, email, document, or form field, hoping an AI system will read them and follow them as if they came from the user or developer.
A common example is an email assistant that summarizes incoming messages. A malicious email might contain hidden text telling the assistant to forward the user's recent invoices to an outside address. If the assistant can send email and has no safeguards, it may comply. Prompt injection is especially serious for AI agents that can browse, send messages, or change records. Defenses include limiting tool permissions, separating trusted instructions from untrusted content, requiring confirmation for sensitive actions, and monitoring for unusual behavior.
Rules, filters, and checks placed around an AI system to keep its inputs, outputs, and actions within safe and acceptable limits.
Agentic AIAI systems designed to pursue goals with some autonomy, planning steps, using tools, and acting on results rather than only answering prompts.
JailbreakBypassing the restrictions on a device or an AI model, either to remove manufacturer limits or to make a model ignore its safety rules.
System PromptBackground instructions given to a language model by the developer that set its role, tone, rules, and limits for every conversation.
AI is reshaping tech roles, but strong fundamentals still win jobs. Here are the skills employers are prioritising in 2026 and a practical plan to build them.
A no-code walkthrough for building an internal AI assistant with instructions, knowledge files and guardrails, including what OpenAI's switch from GPTs to plugins means for you.
AI agents have moved from demos to daily operations. Here is how agentic workflows actually work, where they pay off, and how to roll them out without losing control.