Why Image and PDF Inputs Create a Distinct Attack Surface
Text-based prompt injection defenses generally assume that malicious instructions arrive as part of the visible user prompt, where they can be inspected, filtered, or flagged before reaching the model. Multimodal agents break this assumption. When an agent ingests an image or a PDF, the content that ultimately reaches the model's context window is not necessarily the same content a human reviewer sees when opening the file.
OWASP's Top 10 for LLM Applications formally distinguishes direct prompt injection, where a user authors malicious text, from indirect prompt injection, where instructions are embedded in external content such as documents, websites, or files the model processes. Image and PDF inputs are a primary vector for indirect injection because the ingestion pipeline itself, not the end user, introduces the untrusted content into the model's context.
This matters because most text-based input sanitization operates on the assumption that the visible prompt is the only prompt. Multimodal pipelines have multiple extraction and interpretation stages, each capable of introducing content that was never visible to, or authored by, the person submitting the file.
How Images Introduce Injection Risk
Multimodal agents typically process images through one of two paths: a vision encoder that interprets pixel content directly, or an OCR pipeline that converts visual text into machine-readable text. Once OCR extracts text from an image, that text is commonly treated by the downstream LLM as legitimate prompt context, with the same weight as instructions from a trusted system prompt or a verified user.
This creates two separate attack surfaces rather than one. An attacker can target the OCR/text-extraction layer, embedding instructions that render as normal text once extracted, or target the vision encoder itself, since vision-language models interpret pixel content independently of any extracted text. Effective defense requires validation at both stages, not a single filtering point applied after extraction.
What is confirmed versus theoretical
NIST's adversarial machine learning taxonomy describes evasion and poisoning techniques that manipulate input data to alter model behavior, providing a foundational framework applicable to image-based manipulation. However, specific production techniques such as steganographic text embedding or adversarial patches against deployed multimodal LLM agents were not confirmed in the primary standards and framework documentation reviewed for this guide. These remain plausible extensions of general adversarial ML research rather than confirmed, named incidents.
How PDF Structure Gets Exploited
PDF files are more structurally complex than most security teams assume. The ISO 32000-2 specification permits multiple content layers, embedded metadata (through XMP and document info dictionaries), optional content groups sometimes referred to as layers, and embedded or attached files. These are legitimate document features used for versioning, accessibility, and workflow purposes, but they can be repurposed to hide content that never appears in a standard rendered view.
The operational risk is that PDF ingestion pipelines commonly extract text using parsing libraries that read hidden, layered, or metadata text regardless of whether that text is visible in a standard PDF viewer. A document that looks unremarkable to a human reviewer can still deliver instructions to an agent's context window through a metadata field or a suppressed content layer.
As with image-based techniques, this guide draws a clear line between what is structurally possible and what has been independently confirmed. The PDF specification's technical capabilities support this attack surface, but no primary source reviewed here documents a confirmed real-world AI agent exploitation case within the past twelve months. Security teams should treat this as a credible, specification-supported risk rather than an unverified claim, while avoiding overstatement of prevalence.
Why Downstream Tool-Call Permissions Change the Risk Profile
The distinguishing factor between multimodal prompt injection and text-only prompt injection is not the injection technique itself but what happens after the injected instruction reaches the model. OWASP's LLM Top 10 separates Excessive Agency as its own risk category, noting that agents with broad tool-call permissions and insufficient output validation face materially greater consequences from a successful injection than a pure chat interface would.
An agent that only generates text in response to an injected instruction produces a bad response. An agent that can invoke file access, call external APIs, or execute code in response to that same injected instruction can exfiltrate data, take unauthorized action, or violate policy. Multimodal ingestion increases the number of paths through which an injected instruction can reach that agent, without necessarily changing the permissioning model applied to the agent's tool calls.
This is the core architectural gap enterprises need to close: input validation reduces the likelihood an injected instruction reaches the model, but it does not eliminate the need for permissioning controls on what the agent is allowed to do once an instruction, injected or legitimate, is in its context.
Implementation Priorities for Security Teams
- Treat all extracted text from images and PDFs as untrusted input requiring the same scrutiny as external user-submitted text, not as trusted system context.
- Classify file types and structural features, such as PDF optional content groups and embedded files, before documents reach extraction pipelines.
- Red-team OCR and vision-encoder outputs against known injection patterns as part of pre-deployment testing, using OWASP's indirect injection guidance as a baseline.
- Extend existing DLP and metadata-scrubbing controls already used for document compliance to cover AI ingestion pipelines rather than building parallel tooling.
- Monitor for anomalous tool-call sequences immediately following multimodal ingestion events to catch unauthorized agent actions early.
- Require human-in-the-loop approval for high-impact actions, such as data access or external API calls, triggered by multimodal-derived agent output.
Governance and Evaluation Considerations
NIST's AI Risk Management Framework Generative AI Profile provides a governance structure for identifying and managing prompt injection risk within a broader AI risk management program, but it does not mandate specific multimodal controls. Similarly, OWASP's LLM Top 10 functions as an informal industry reference rather than a binding regulatory standard. Enterprises should treat both as a baseline for risk identification, not as a compliance guarantee that multimodal ingestion is secure.
Given the absence of a unified, multimodal-specific technical standard, organizations enabling image and PDF ingestion in production agents should document risk acceptance decisions explicitly, particularly where injection risks remain unresolved. Cross-functional accountability across security, AI/ML engineering, and legal or compliance functions is advisable when approving the scope of tool-call permissions granted to multimodal agents, given the potential for data exfiltration or policy violation if an injection succeeds.
When evaluating a vendor's multimodal agent architecture, ask whether untrusted extracted content is segregated from trusted instructions, what validation occurs at the OCR and vision-encoding stages, what permissioning and approval controls exist for downstream tool calls, whether extracted text and metadata are logged separately for incident investigation, and whether the vendor has conducted red-team testing specifically against indirect or multimodal injection.
Where Runtime Governance Fits
Input validation and pipeline-stage sanitization reduce the likelihood that an injected instruction reaches an agent's context, but they cannot fully close the risk on their own, since new extraction and encoding techniques will continue to surface. The control that determines actual blast radius is what the agent is permitted to do once an instruction, injected or otherwise, is in play.
This is where runtime governance becomes the necessary complement to input-layer defenses. Enforcing least-privilege permissions on agent identity, requiring tool-approval workflows for high-impact actions, and maintaining audit logging of agent behavior after multimodal ingestion address the consequence side of the risk rather than only the detection side. Trussed AI provides runtime governance and security for enterprise AI agents, including agent permissions, tool approval workflows, and audit logging, applied at the point where an agent attempts to act, independent of how the underlying instruction reached its context.