Last updated: October 7, 2026

TL;DR
  • Prompt injection against robots uses text, speech or images to make the AI model that steers a robot ignore its real task and follow an attacker's instruction instead.
  • Peer-reviewed research has shown it on real hardware, from jailbroken robot dogs to a robot arm misled by a printed note, with success rates from 5% to 100% depending on the model and attack.
  • Simple defenses such as masking text in camera images help a lot, but the safest design keeps hard safety limits outside the AI model.

Prompt injection is the top-ranked risk in the OWASP Top 10 for LLM applications, and it does not stay on the screen when the model drives a body. Humanoids increasingly run a vision-language-action (VLA) model or a vision-language model (VLM) planner that turns spoken words and camera images into motion. If an attacker can get words into what that model reads, they may be able to change what the robot does. This guide explains how prompt injection against robot VLA models works, what the research shows so far, and which defenses hold up.

What is prompt injection in robotics?

OWASP defines prompt injection as inputs that "alter the LLM's behavior or output in unintended ways," and splits it into two kinds (OWASP LLM01):

  • Direct injection: the person talking to the model crafts the input, for example a spoken command phrased to slip past the robot's safety instructions. In robotics this is usually called a jailbreak.
  • Indirect injection: the instruction hides in content the model reads from somewhere else. For a robot, "somewhere else" includes the physical world: a sign, a label, a sheet of paper or a screen in the camera's view.

The robotics twist is the output. A chatbot that is tricked writes bad text. A robot that is tricked can move, grab, drive or open something. It also widens the attack surface: anyone who can place an object in front of the camera can, in principle, send the model a message.

How does prompt injection reach a robot's VLA model?

A modern humanoid stack has several entry points where untrusted language can reach the model that plans actions:

Entry point Example Type
Voice or chat interface A user phrases a request to bypass safety rules Direct (jailbreak)
Text in the camera view A printed note, sign or label with instructions Indirect (physical)
Images crafted for the model Adversarial patches or hidden text humans don't notice Indirect (visual)
Data the robot retrieves Task lists, web pages, emails or work orders it reads Indirect (digital)
Fleet or cloud channels A compromised operator console or task queue Indirect (supply chain)

The AI layer is also only one part of the humanoid attack surface. The language interface can also be a door to classic exploits: one of the Unitree G1 flaws covered in our CVE-2026-76639 explainer reached root code execution through the robot's AI chatbot service.

What has research shown so far?

All of the results below come from published academic work. As of this update we know of no confirmed prompt injection attack against a deployed commercial humanoid in the wild.

Jailbreaking LLM-controlled robots (2024). Researchers at the University of Pennsylvania built RoboPAIR, which automatically crafts jailbreak prompts for robots. They tested it on the NVIDIA Dolphins self-driving model, a Clearpath Jackal ground robot running GPT-4o and a Unitree Go2 robot dog, and report that it "frequently" reached 100% attack success (Robey et al., project site).

Text attacks on VLA models (2025). A study of common VLA models found that a single textual attack at the start of a task could give "full reachability of the action space," and that the effect often persisted over long tasks. The authors note that robotic attacks "do not have to be semantically linked to notions of harm," which makes them harder for content filters to catch (Jones et al.).

Command hijacking with signs (2025-2026). CHAI, presented at IEEE SaTML, embedded deceptive instructions in visual input such as misleading signs. It was tested on drone emergency landing, autonomous driving, aerial tracking and a real robotic vehicle, and the authors report it consistently outperformed earlier attacks (Burbano et al.).

"Hijacking robots with a piece of paper" (August 2026). The most systematic test so far placed printed instructions in the workspace of a Kinova robot arm doing a sorting task, steered by three vision-language models (Samarakoon et al.):

Model Overall attack success
GPT-4o 27.0%
Gemini 2.5 Flash 29.4%
Qwen3-VL-32B 5.0%

Notes that impersonated an authority were the most effective, reaching 56.7% to 80.0% success on the closed-source models, while indirect signage reached 0% on all three. The models acknowledged the injected text in 99.9% of successful attacks, which the authors read as the models deliberately deferring to it rather than being fooled.

Adversarial images (2024-2026). Separate work shows that small patches in the camera view can derail VLA models and even transfer between models; see our VLA glossary entry for details.

Why prompt injection matters for humanoid security

Analysis: three features make this risk sharper for humanoids than for chatbots.

  1. The input channel is the room. A humanoid in a warehouse, hospital or home reads its surroundings all day. Anyone who can tape a sheet of paper to a shelf can address the model.
  2. The output is physical. Even a 5% success rate is too high if the action is opening a door, handing over an item or moving fast near people.
  3. Filters miss it. Research shows harmful robot actions can be triggered by text that looks harmless, so blocklists built for chatbots are a weak fit.

How do you defend robots against prompt injection?

No single control solves prompt injection, in robots or anywhere else. Use layers:

  • Keep hard safety limits outside the model. Speed, force, workspace boundaries and an independent emergency stop should be enforced by safety-rated controllers the AI cannot override.
  • Treat everything the camera reads as untrusted. In the 2026 study, masking detected text before the model saw the image blocked 100% of the attacks; telling the model to ignore in-scene text was 75% to 100% effective, and a second verification query was 85% to 100% effective (Samarakoon et al.). Weigh the trade-off for tasks that need to read real labels.
  • Least privilege for the planner. Give the model a short list of allowed skills instead of open-ended motion, in line with OWASP's mitigations.
  • Human approval for high-risk actions, such as opening doors, handling hazardous items or leaving a zone.
  • Authenticate who can give commands. Voice and chat interfaces should know who is speaking and what they are allowed to ask.
  • Log prompts, images and actions so you can investigate odd behavior.
  • Red-team before deployment, including physical tests with printed instructions in the workspace.
  • Patch the plumbing too. Language interfaces have led to classic code execution bugs; track them in our Vulnerability Tracker.

Our free Humanoid Robot Security Checklist turns these into a deployment review, and the Directory lists AI security vendors that address prompt injection.

FAQ

Can a humanoid robot be prompt injected? Research shows robots steered by language and vision-language models can be manipulated by crafted text, speech and printed instructions. We know of no confirmed real-world attack on a deployed commercial humanoid as of this update.

Is prompt injection the same as a jailbreak? They overlap. A jailbreak usually means the user directly tricks the model past its safety rules. Prompt injection also covers indirect attacks, where instructions hide in content the robot reads, such as a sign in the camera view.

What is the best defense against robot prompt injection? Layered controls: hard safety limits outside the AI model, treating in-scene text as untrusted, limiting what the model is allowed to do, and human approval for risky actions.

Do content filters stop prompt injection in robots? Not reliably. Research on VLA models found that harmful actions can be triggered by text that is not obviously about harm, so filters built for chatbots miss many attacks.

Get the weekly Humanoid Threats Brief → Subscribe