Glossary
Prompt injection
An attack where instructions hidden in content the model reads override the behaviour you intended.
If your system reads anything a user or third party supplies, an email, a web page, an uploaded document. That content can carry instructions. The model has no reliable way to distinguish data from instruction.
Defences are architectural rather than textual: least-privilege tools, approval gates on consequential actions, treating retrieved content as untrusted, and output filtering. Telling the model to ignore injected instructions is not a defence.
The most reliable mitigation is architectural separation: an agent that reads untrusted content should not be the same agent that holds write permissions. Splitting those roles costs a little design effort and removes the entire class of escalation.
Commonly misunderstood: It is often dismissed as theoretical. It becomes concrete the moment an agent with write permissions reads untrusted input.
Related terms, in context
The concepts you almost always meet alongside prompt injection.
- Red teaming
- Deliberately attacking your own AI system to find failures before users or attackers do.
- AI agent
- A system that plans a sequence of steps, calls tools, and changes state in real systems, rather than only producing text.
- Guardrails
- Constraints on what an AI system may do, enforced in code rather than requested in a prompt.
- Tool calling
- Letting a model invoke defined functions with structured arguments, so it can act rather than only describe.
Where this shows up in our work
Prompt injection is not an abstraction for us. It is a decision we make on live projects. It shows up most directly in ai evaluation & red teaming, ai governance & compliance, where getting it wrong has a cost someone can measure.
If you are evaluating a vendor on this, the useful question is not whether they can define the term. It is what they measure, what they would refuse to do, and what happens in their system when the assumption behind prompt injection stops holding.
Questions
What is Prompt injection?
An attack where instructions hidden in content the model reads override the behaviour you intended.
What do people get wrong about prompt injection?
It is often dismissed as theoretical. It becomes concrete the moment an agent with write permissions reads untrusted input.
Does Orqent Labs build this?
Yes, AI Evaluation & Red Teaming and AI Governance & Compliance. We work across India, covering all 19,238 PIN codes remotely.
Building something that involves prompt injection?
We will tell you honestly whether it is the right approach for your problem.
Or email bd@dtrasglobal.com · call +91 74118 77878