1. Introduction

Large language model (LLM) agents are systems that plan multi-step tasks, call external tools, browse the web,

and act on retrieved information with minimal human supervision which have moved from research prototypes

to production deployments across coding assistants, customer support, and enterprise workflow automation

within roughly two years. This shift toward autonomy has been accompanied by a corresponding shift in the

attack surface: rather than attacking a model directly through a user-typed prompt, adversaries increasingly plant

malicious instructions inside content the agent is expected to read as part of its normal operation — a web page,

a retrieved document, a tool's output, an email, a code comment. This class of attack, known as indirect prompt

injection, exploits the fact that LLM agents cannot reliably distinguish between trusted instructions from their

operator and untrusted data encountered during execution. Industry threat reporting for 2026 identifies agentic

systems as the fastest-growing category of AI-related security incident, and standards bodies including OWASP

and the U.S. National Institute of Standards and Technology (NIST) have, within the past several months,

published dedicated guidance acknowledging that existing defenses were not designed with autonomous,

context-consuming agents in mind. against indirect prompt injection specifically in the setting where an agent's use of external context is legitimate

and necessary — a setting current defenses handle poorly.

2. Background

A typical loop for an LLM agent is: Take in a task, reason about what to do next, then call a tool (such as a

search engine, code interpreter, email client, file system, etc), and then observe its output, then repeat until the

task is complete. The model is provided with both the operator's original instruction and the content returned by

the tools used in it, and as current LLMs lack a structural distinction between “instruction” and “data”, any text

that reaches the “instruction” channel could be interpreted as a new instruction. The indirect prompt injection

is all about this: if the attacker can't directly interact with the model, they can still manipulate the model by

inserting their prompt at a location where the agent draws them.

This was mainly considered as a filtering problem; identify suspicious text and remove it before the model.

More recent approaches to architecture try to deliver more guarantees: an explicit hierarchy of instructions

(Wallace et al., 2024), having an isolated “quarantined” LLM process untrusted data while a separate, privileged

LLM takes actions (Kim, Choi & Lee, 2025), or restructuring the agent so it has one “quarantined” LLM for

data processing and another privileged LLM for action-taking (Debenedetti et al., 2025). Evaluation has also

developed over these defenses: benchmark environments like AgentDojo (Debenedetti et al., 2024), InjecAgent

(Zhan et al., 2024), and WASP (Evtimov et al., 2026) now make it possible to assess attacks and defenses under

tool-assisted agent scenarios in controlled settings.

Nevertheless, although thus far impressive, a 2026 systematic review of the field revealed that existing defenses

and benchmarks have a structural blind spot: they are evaluated on tasks where the agent's ideal behavior is to

ignore untrusted content, but not on the more realistic and more common scenario, where the agent is authorized

indeed required to act on the content it retrieves at runtime (Wang et al., 2026). In such benchmarks, merely

suppressing an agent's willingness to use retrieved context will result in a good score, and violate the tasks

agents are meant to perform.

3. Problem Statement

The core challenge that this research tackles is the absence of defenses against indirect prompt injection that

have been demonstrated to withstand the most important scenario for actual deployment: context-dependent

agentic tasks, where the agent is required to read and follow instructions from external content that it does not

trust to be part of its task. In existing evaluations, no single defense is simultaneously high in both

trustworthiness (resistance to agent injection) and task utility (ability to complete the legitimate task) and low

in latency (time required to complete the task) (Wang et al., 2026). With this is an agent that does everything

1by consuming untrusted information from the outside world, such as a web -browsing assistant, an agent that

retrieves material from a remote source or a document processing pipeline, and no effort has been made to

validate it an agent that is common today but has been left without a validated method to achieve. Thus there is

a clear, current gap between the security current defenses claim and the security which is available to agents

who can't avoid using untrusted context.

4. Research Objectives

1. Characterize how the effectiveness of representative existing prompt-injection defenses changes on

context-dependent agentic tasks, relative to their reported performance on standard, context-suppressible

benchmarks.

2. Identify the mechanism(s) by which context-dependent tasks defeat or degrade current defenses —

whether through an inability to distinguish legitimate from illegitimate context use, utility/latency trade-

offs, or benchmark design gaps.

3. Design and implement a context-aware defense mechanism that explicitly attempts to distinguish an

agent's legitimate use of retrieved information from an attacker's attempt to hijack the agent's behaviour

through that same information.

4. Evaluate the proposed defense against existing baselines on trustworthiness, task utility, and latency, using

existing and, where necessary, extended benchmark tasks.
5. Research Questions

1. How much does the effectiveness of current indirect-prompt-injection defenses degrade when evaluated

on context-dependent agentic tasks compared to context-suppressible tasks?

2. What distinguishes cases where an agent's use of retrieved context is legitimate from cases reflecting

successful injection, in terms of features observable at inference time (e.g., instruction–content similarity,

action consequence, source provenance)?

3. Can a defense mechanism be designed that uses such features to permit legitimate context-dependent

reasoning while blocking or flagging injected instructions, without a large utility or latency penalty?

4. How does the proposed defense's trustworthiness–utility–latency trade-off compare to existing baseline

defenses across a range of task types and attack strategies?