EN
Sign inGet started
Glossary

Context Window Manipulation

What is Context Window Manipulation?

Context window manipulation refers to attacks that deliberately influence the text a language model processes at one time, for instance through hidden instructions in documents, fabricated earlier parts of a conversation or large amounts of filler that push the original rules into the background. Protection comes from clearly separating trusted instructions from external content, length limits, input checks and granting agents as few permissions as possible.

How KI.ordo helps

KI.ordo's “AI firewall: check input” template can be placed in front of an agent as a sub-flow to filter out suspicious inputs beforehand.

Learn more →: Context Window Manipulation

Matching templates

Topics

  • Security
  • Prompt injection
  • LLM

Try KI.ordo

Agents, flows and templates in one workspace – free 24-hour demo.

Start demo
← All terms

Related terms

Indirect Prompt Injection

In an indirect prompt injection, manipulative instructions are not in the user's input but hidden in content an AI system processes along the way, such as web pages, emails or documents. When an agent reads such content, it may mistake hidden commands for genuine tasks and, for example, disclose data. Safeguards include treating external content strictly as data, limiting permissions, screening inputs and requiring human approval for critical actions.

  • Security
  • Prompt injection
  • +1
Template

AI Firewall

An AI firewall is a protective layer between users, data sources and an AI model. It screens inputs for manipulation attempts such as prompt injection or jailbreaks and checks outputs, for example for confidential data or unwanted content, before passing them on. It may use rules, classification models or a second language model. An AI firewall reduces risk but does not offer complete protection and replaces neither limited permissions nor testing.

  • Security
  • Prompt injection
  • +1
Template

Token Smuggling

Token smuggling is a form of jailbreaking in which unwanted instructions are disguised, for instance split into fragments, re-encoded or written with look-alike characters, so that simple keyword filters miss them while the language model still grasps their meaning. Protection includes normalising inputs before they are assessed, content-based rather than purely word-based checks, reviewing outputs and granting AI agents as few permissions as possible.

  • Security
  • Jailbreaking
  • +1
Template

Jailbreaking

Jailbreaking refers to attempts to bypass an AI model's safety rules through cleverly worded inputs, so that it produces content or takes actions it is meant to refuse. For organisations this is a risk, because a chatbot could, for example, reveal confidential instructions or respond inappropriately. Protective measures include clear system instructions, input and output screening, limited permissions, logging and regular security testing (red teaming).

  • Security
  • Misuse prevention
  • +1
Template