Have an AI product going live?
Let's Talk

Beyond the Prompt: Physical-World AI Injection Attacks

How physical devices and everyday workflows can carry persistent prompt injections into AI-connected systems.

In a previous post, we presented prompt injection research that focused on digital content crossing the cyber-physical barrier. The resulting malicious instructions embedded in websites, documents, repositories, and API responses were able to influence AI systems when the content was retrieved and interpreted as part of a legitimate task. These attacks typically exploit a familiar trust boundary where an AI system receives information from a source it should treat as data, but the model interprets some of that information as instructions.

Our recent research focus has explored how enterprise systems routinely collect information through scanners, access-control devices, wireless peripherals, and other physical interfaces. These inputs often pass through several applications before reaching databases, logs, or reporting systems.

When AI is introduced downstream, these existing data flows become potential prompt injection delivery channels. The below research focuses on understanding how attacker-controlled instructions move through these environments, where they persist, and what prevents them from reaching an AI system. Importantly, successfully delivering a payload is a separate problem from persuading a model to follow it.

‍

Expanding the Prompt Injection Attack Surface

Indirect prompt injection occurs when an attacker places instructions in content that an AI system encounters while carrying out an authorized task. The attacker does not control the user's original request or the system's instructions. Instead, they control some portion of the information the model is expected to process.

Physical-world prompt injection extends this technique by moving the initial delivery point into an everyday device or operational process, which Mindgard’s research defines as a a five-stage attack chain: 

‍

    ‍

    • INJECT: The INJECT stage introduces attacker-controlled content through an input channel such as a QR code, device identifier, or other physical interface. The objective is to have that content accepted by an existing system as ordinary system data.
    • STORE: During STORE, the content is written to a database, log, asset record, or similar repository. At this point, the attacker-controlled text may become indistinguishable from information collected through legitimate business activity unless its origin is preserved.
    • DORMANT: The DORMANT stage separates initial delivery from the eventual AI interaction. The payload can remain stored without producing an immediate effect.
    • TRIGGER: TRIGGER occurs when an AI-connected workflow retrieves the stored information. This might happen during compliance reporting, inventory analysis, security investigation, or another automated task.
    • EXECUTE: Finally, EXECUTE represents the attempted behavioral outcome. The attacker wants the model to interpret the embedded text as an instruction and act on it, rather than processing it solely as data.

    These stages describe the intended attack path, not a guarantee that every stage will succeed. The payload may be rejected, modified, ignored, or never encountered by an AI system.

    ‍

    Understanding Blind Prompt Injection

    One of the defining characteristics of physical-world prompt injection is the limited visibility available to the attacker. In a conventional prompt injection test, a researcher may control a document or webpage and immediately observe how the model responds. That feedback helps establish whether the payload reached the model and whether it influenced the output.

    Physical delivery often removes this feedback loop. An attacker could introduce malicious content through a device or label without knowing whether the receiving system accepted the complete payload. Even if the content is stored successfully, the attacker may not know when an AI workflow will retrieve it.

    This is a form of blind prompt injection where uncertainty exists at several points in the execution path. Hardware may truncate or reject the input. An application may sanitize or transform it. A database may retain only part of the original content. A downstream AI workflow may never retrieve the relevant field. Even if the original text reaches the model, the attempted instruction may have no effect.

    For researchers, this makes it important to separate delivery validation from model-behavior evaluation. Evidence that a payload was stored is useful, but it does not establish that the attack reached its intended final stage.

    ‍

    Tracing the Attack from Physical Input to AI Processing

    To investigate this attack path, Mindgard examined a workflow involving Snipe-IT, an open-source IT asset management platform used to track hardware and associated inventory information.

    ‍

    Mindgard’s research demonstrated that a printed QR label could act as the initial carrier for attacker-controlled content. In normal operation, an asset label allows someone to identify equipment or interact with the corresponding record through a scanning workflow. The scanner acts as a bridge between the physical object and the digital asset management system.

    The initial step involved preparing a QR label containing attacker-controlled text. When processed through the scanning workflow, the encoded content could enter the asset management application and be preserved within its records.

    ‍

    ‍

    The next stage involved a downstream AI-generated compliance workflow consuming information from those records and an attempt to compromise the relevant trust boundary. The goal was simple: provide the asset management system malicious information that appears to be valid, but is interpreted by AI to generate a narrative or assessment.

    If attacker-controlled text is included in the model's context, the model may encounter instructions that were never provided by the legitimate user.

    ‍

    Investigating Physical Attack Vectors

    This proof-of-concept used QR a carrier, but the underlying technique is not limited to barcodes and printed labels. Many physical devices expose text fields or identifiers that are collected automatically by enterprise software. Whether these fields are suitable for prompt injection depends on their format, length, character restrictions, and how receiving systems process them.

    Conference badges provide one potential entry point. Information encoded in a badge or associated identifier may be collected through event management, visitor registration, or access workflows. The relevant question is whether attacker-controlled text survives those processes and reaches a system that later supplies context to AI.

    USB device descriptors offer another area for investigation. Connected devices can expose manufacturer, product, and serial-number strings to the host operating system. Depending on how these values are collected, they may appear in EDR audit trails or diagnostic records.

    Bluetooth Low Energy devices can advertise names that nearby systems discover and display. These names may subsequently be recorded by monitoring, inventory, or device management applications.

    Certificate fields also contain potentially attacker-controlled text, including subject names and other metadata. Applications that inventory certificates or summarize their contents may propagate these values into reports or AI-generated analysis.

    In each case, the existence of a text field is only the starting point. A researcher must establish whether the attacker can control the value, whether the receiving system accepts it, how it is transformed, and whether the resulting information is available to an AI workflow.

    ‍

    Understanding Where Attacks Fail

    Not every physical input channel provides a viable route for delivering prompt injection payloads. Our research also included unsuccessful experiments involving access-control readers. These negative results are useful because they illustrate how the constraints of physical systems can interrupt the attack chain before an AI system becomes involved.

    Unlike general-purpose text inputs, many access-control devices operate using defined credential formats and communication protocols. A reader may accept only specific identifiers, fixed-length values, or data structured according to a particular standard. Even when an attacker can influence the transmitted information, that does not mean arbitrary natural-language text can pass through the system.

    ‍

    ‍

    Additional transformations can occur between the reader and the receiving application. Input may be converted into numeric identifiers, normalized, truncated, or discarded before reaching persistent storage. These behaviors can prevent a prompt injection payload from surviving in a form that a language model could later interpret.

    The unsuccessful reader experiments therefore help distinguish a plausible physical entry point from a demonstrated end-to-end delivery mechanism. For researchers, understanding these failures is as valuable as demonstrating a successful delivery path. The important measurement is not simply whether a device accepts attacker input, but whether the relevant content survives each processing boundary intact.

    ‍

    Evaluating Model-Dependent Execution

    Once malicious content reaches an AI system, the remaining outcome depends on how the model and surrounding application handle untrusted information. Successful injection transport does not guarantee successful prompt injection execution. A model may treat the embedded instruction as irrelevant source text, decline to follow it, or produce an output unaffected by the payload. Application controls may also prevent an attempted instruction from producing a consequential action.

    Conversely, unsafe interpretation may allow untrusted content to influence a generated response or an agent's subsequent behavior. This distinction is central to evaluating the research accurately. The physical delivery mechanism determines whether the payload reaches the AI workflow. The model and application determine what happens next.

    ‍

    Identifying the Trust Boundary Failure

    The underlying security concern is the movement of information between systems with different trust assumptions. A scanner, hardware interface, or device management service may be designed to collect data without assigning meaning to its contents. That data is then preserved by a database for later retrieval, which in turn may introduce malicious commands to an AI system that processes the retrieved text as natural language, combining it with user instructions, application context, and other records to complete a task.

    The trust boundary fails when content originating from an untrusted physical input is interpreted as authoritative guidance for the AI system. This does not require the initial application to malfunction, every preceding component may be operating according to its intended design. The problem emerges from the interaction between those components and the downstream AI workflow.

    For security researchers, identifying this boundary requires examining the complete data path. Testing only the model or only the physical device will miss important conditions that determine whether an attack is viable. This includes the original input format, storage behavior, retrieval logic, prompt construction, and permissions available to the AI application. A complete assessment therefore needs to establish both the integrity of the payload's delivery path and the authority assigned to that content when it reaches the model.

    ‍

    Introducing the Research Tooling

    To support investigation into these delivery mechanisms, our team has introduced two open-source resources: Prompt Injection Studio and the Flipper Zero Field Kit.

    • Prompt Injection Studio is intended to support the preparation and investigation of prompt injection payloads across different delivery scenarios.
    • Flipper Zero Field Kit focuses on the physical side of the research, providing resources for exploring supported hardware-oriented input channels.

    Together, these resources are intended to help researchers investigate how malicious text can enter operational systems, how that text is represented after ingestion, and where the delivery path may break. The objective is to make the transport layer of prompt injection easier to examine independently of downstream model behavior.

    ‍

    Closing Perspective

    Physical-world prompt injection extends an established AI security problem into systems that were not originally designed around language-model interpretation, resulting in previously trusted systems providing routes for attacker-controlled content to reach AI workflows through legitimate business processes.

    The [un]prompted.au demonstration illustrates how physical delivery and persistence can be investigated. The unsuccessful access-control experiments show why the limitations of individual protocols and data transformations also matter. For researchers, a meaningful assessment must cover the entire chain, from the physical input to the stored record, retrieval process, and eventual AI interaction.

    As organizations connect more operational systems to AI, understanding these paths will become an increasingly important part of testing how models encounter and interpret information from the outside world.

    ‍

    ATTACKER-ALIGNED AI SECURITY

    Talk to Our AI Security Experts

    We found vulnerabilities in AI from OpenAI, Meta and AWS. See how Mindgard finds and proves the weakness attackers could exploit in yours.

    ✖