Sovra Recon·Glossary · Prompt Injection

Prompt Injection: Definition, Examples, and Protection

By Solan Barlet, co-founder of Sovra Recon
Updated October 29, 2026

In short

Prompt injection involves inserting instructions into the text read by a language model to manipulate its behavior: causing it to ignore its instructions, reveal information, or use its tools on behalf of the attacker. It can be direct (typed into the conversation) or indirect (hidden in an email, document, or web page that the AI accesses). This is the first risk listed in the OWASP Top 10 for LLM Applications.

Direct or Indirect

  • Direct: The user writes instructions themselves that circumvent the system’s guidelines.
  • Indirect: The instructions are hidden in content that the AI processes on behalf of someone else, such as invisible text on a web page summarized by an assistant, or in an email read by an agent who has access to the inbox.

Indirect injection is the most dangerous: the victim didn’t type anything; they simply asked their assistant to read some malicious content.

Mitigating the Risk

  • Grant the AI the minimum necessary permissions and data access.
  • Require human confirmation for any sensitive action.
  • Clearly separate instructions from external content, and treat the latter as untrustworthy.
  • Test regularly using realistic injection scenarios.

Frequently asked questions

Is a filter enough to prevent prompt injection?

No. Filters reduce the simplest attacks but can be circumvented. Protection relies primarily on limiting what the AI can read and do.

Sources

Find out where an attacker would get in.

Our AI agents actually attack your scope, fully automated, and hand you the proof of every vulnerability.