What is prompt injection?
Prompt injection is an attack in which text supplied to a language model changes its behaviour in ways its operator did not intend. Direct prompt injection arrives in the input a user types. Indirect prompt injection arrives in content the model reads while it works, such as an email, a web page, a document or a tool result.
Also called direct prompt injection, indirect prompt injection. Last reviewed .
Why it matters
OWASP ranks it first: LLM01:2025 Prompt Injection. NIST's adversarial machine learning taxonomy treats direct and indirect prompt injection as distinct attack classes, because the indirect kind needs no access to the system at all, only a way to put text where the model will read it.
CVE-2025-32711, known as EchoLeak and published on 11 June 2025, is the reference case. Microsoft describes it as "AI command injection in M365 Copilot" that "allows an unauthorized attacker to disclose information over a network", and scored it 9.3.
When the model is an agent, a successful injection is no longer a wrong answer. It is whatever the agent is able to do with its tools and credentials.
Related terms
- Tool poisoning
- Tool poisoning is indirect prompt injection delivered through a tool description or result.
- Excessive agency
- Excessive agency sets the cost of a successful injection.
- Confused deputy
- An injected agent acting with its own broad access is a confused deputy.
- AI red teaming
- AI red teaming is how you find out which injections work against your agents.
How Guard0 handles it
Guard0 red-teams agents with prompt injection delivered the way attackers deliver it, through tickets, documents and tool results. Each finding reaches the owner of the agent with the payload and every tool call that followed, and the same payload runs again after the fix.
Sources
- LLM01:2025 Prompt Injection. OWASP Gen AI Security Project.
- CVE-2025-32711. NVD.
- CVE-2025-32711: M365 Copilot Information Disclosure Vulnerability. Microsoft Security Response Center.
- Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST AI 100-2 E2025). NIST.