Direct prompt injection targets the model through an explicit malicious instruction given to the agent. Indirect prompt injection hides the malicious instruction inside external content such as a webpage, document, or email that the agent later reads. For security teams, the distinction matters because indirect attacks exploit trust in third-party content and are harder to spot through normal user oversight.
How direct and indirect prompt injection differ in practice
Direct prompt injection is the more obvious form: the attacker places a malicious instruction directly into text the agent is asked to follow. indirect prompt injection works one layer away, by embedding that instruction in content the agent later consumes, such as a webpage, document, ticket, or email. The security difference is trust boundary, not syntax.
That distinction matters because an agent may treat external content as data while still using it as instruction material. In a direct attack, the malicious intent is usually visible at the point of interaction. In an indirect attack, the harmful instruction can hide inside ordinary content, so the agent’s decision path becomes the weak point.
Why indirect prompt injection is harder to spot and contain
Direct prompt injection often fails when users or automated filters notice the hostile instruction. Indirect prompt injection is more difficult because the malicious payload can travel through normal workflows, then activate only when the agent retrieves or reads the content. That makes it a trust abuse problem as much as a content problem.
This is why browser-driven agents, document-reading assistants, and email-connected agents deserve special scrutiny. The agent is not being “hacked” by the file format alone; it is being manipulated through a trusted input path. The attack succeeds when the system fails to separate retrieved content from actionable instructions.
For a useful security mental model, treat direct prompt injection as a hostile command attempt and indirect prompt injection as instruction smuggling through an untrusted source. The mitigation emphasis shifts from only policing user prompts to controlling what the agent can read, how it interprets it, and which external sources are allowed to influence action.
What defenders should evaluate first
Start by mapping where the agent can ingest outside content and whether that content can influence tool use, memory, or delegation. The highest-risk cases are agents that can read arbitrary web pages, process untrusted files, or act on email and chat content without a strong confirmation step. Those paths expand the attack surface far beyond the original prompt.
Next, test whether the agent can be induced to follow instructions that are not clearly separated from content. A safe design should make the agent resistant to instruction buried in retrieved text, and should require explicit policy checks before high-impact actions. If a simple content fetch can change behaviour, the boundary is too loose.
Useful controls include source allowlisting, content sanitisation, per-action approval gates, and limiting what retrieved text can influence. The browser-and-computer-use pattern is especially sensitive here because a page can contain both legitimate content and malicious instructions, and the agent may not reliably distinguish them without guardrails. Browser and Computer-Use Agent Security Guide
Risk and Threat Considerations
Indirect prompt injection is usually the more dangerous variant because it abuses a trusted content pipeline and can bypass normal human review. It becomes especially risky when the agent can act on behalf of a user or invoke tools after reading the compromised content.
Failure mechanism: The attacker hides instructions inside content the agent trusts, then relies on the agent to treat that content as operational input rather than passive information. The payload may trigger when the agent summarises, plans, or executes a tool action.
Impact: The result can be unauthorized actions, data exposure, policy bypass, or manipulation of downstream decisions. In agentic systems, that may extend from misinformation to tool misuse and privilege abuse if the instruction reaches an action boundary. OWASP Agentic AI Top 10 MITRE ATLAS adversarial AI threat matrix
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10, MITRE ATLAS and OWASP API Security Top 10 define the specific risk controls and attack patterns relevant to this topic.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI01 — Agent Goal Hijack | Prompt injection can redirect an agent from its intended goal. |
| ASI02 — Tool Misuse | Injected instructions may cause unsafe tool actions or unwanted execution. | |
| ASI03 — Identity & Privilege Abuse | Injected content can induce actions that exceed intended authority. | |
| Recommendation — Constrain agent goals so untrusted content cannot redirect task execution. Gate tool calls with policy checks before executing untrusted instructions. Bind each action to the minimum authority needed and require explicit approvals. | ||
| MITRE ATLAS | MITRE ATLAS | ATLAS catalogs adversarial AI techniques including prompt injection and manipulation. |
| Recommendation — Map prompt-injection scenarios to adversarial techniques and test detections against them. | ||
| OWASP API Security Top 10 | API6 — Unrestricted Access to Sensitive Business Flows | Injected instructions may push an agent into sensitive business actions through trusted flows. |
| Recommendation — Protect sensitive flows with authorization checks before an agent can complete them. | ||
Practitioner Guidance
What to prioritise: Treat any agent that reads untrusted content as a boundary problem first, not a prompt-quality problem. The key question is whether external text can change tool use or privilege-bearing actions without an explicit policy decision.
What to verify: Confirm that the agent can distinguish content from instructions, that retrieved material cannot silently escalate its own importance, and that high-impact actions require a separate authorization step. If the system cannot explain why an instruction was trusted, you do not yet have a reliable control.
Decision rule: If the agent can both read external content and act on it, assume indirect prompt injection is in scope and constrain the action path before expanding the content corpus. Direct prompt injection is easier to notice; indirect prompt injection is easier to operationalise.
Practitioner takeaway: The real security boundary is not “prompt versus no prompt”, but “trusted instruction versus untrusted content”, and indirect attacks are the ones most likely to cross that boundary unnoticed.
Related resources from NHI Mgmt Group
- What is the difference between prompt injection and pre-task authority in AI agent security?
- What is the difference between prompt injection and excessive privilege in agentic AI?
- What is the difference between prompt injection and traditional injection attacks?
- What is the difference between prompt injection and indirect prompt injection?
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on September 30, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org