> ## Content Index
> Fetch the complete content index at: https://www.humanoidthreats.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Can LLM-Controlled Robots Be Jailbroken?
- URL: https://www.humanoidthreats.com/can-llm-controlled-robots-be-jailbroken/
- Published: 2026-10-21T12:59:09.000Z
- Updated: 2026-11-16T06:58:06.000Z
- Description: Yes. Research shows LLM-controlled robots can be jailbroken into unsafe actions, often with 100% success. See how it works and which defenses help.
- Author: Jonas Weber
- Tags: Q&A

Yes. Peer-reviewed research shows that robots controlled by large language models (LLMs) can be jailbroken: tricked with carefully worded requests into planning and carrying out physical actions their safety rules should block. In tests on real robots, including a commercial robot dog, attacks often succeeded 100% of the time.

*Last updated: October 21, 2026*

## What does it mean to jailbreak an LLM-controlled robot?

A jailbreak is a prompt designed to get an AI model to ignore its safety instructions. In a chatbot, the result is harmful text. In a robot, the LLM's output becomes a plan, code or motion command, so a jailbreak turns words into physical actions. It is the direct form of [prompt injection against robot AI models](https://www.humanoidthreats.com/prompt-injection-robot-vla-models/): the attacker talks to the robot through its normal voice or chat interface, rather than hiding instructions in something the robot sees.

This matters for humanoids because many of them now use an LLM or a [vision-language-action (VLA) model](https://www.humanoidthreats.com/what-is-vision-language-action-model/) as their high-level "brain," the layer that decides what to do next.

## What has research shown so far?

Three studies, in order of publication, show how the problem has developed:

| Study                                                                                                                                                                             | What was tested                                                                            | Key finding                                                                                                              |
| --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------ |
| [BadRobot](https://proceedings.iclr.cc/paper%5Ffiles/paper/2025/hash/5b2fa23e4ef0f7ac6c4f01d7998e6237-Abstract-Conference.html?ref=humanoidthreats.com) (Zhang et al., ICLR 2025) | Embodied LLM frameworks VoxPoser, Code as Policies and ProgPrompt                          | Ordinary voice-style requests could push robot manipulation systems to break safety and ethical constraints              |
| [RoboPAIR](https://arxiv.org/abs/2410.13691?ref=humanoidthreats.com) (Robey et al., University of Pennsylvania)                                                                   | NVIDIA Dolphins driving LLM, Clearpath Jackal with a GPT-4o planner, Unitree Go2 robot dog | Attacks "often" reached 100% success; the authors call the Go2 result the first jailbreak of a deployed commercial robot |
| [Blindfold](https://arxiv.org/abs/2603.01414?ref=humanoidthreats.com) (Huang et al., March 2026)                                                                                  | Embodied AI simulators and a real 6-degree-of-freedom robot arm                            | Instructions that each look harmless can add up to a harmful action, with up to 53% higher success than earlier attacks  |

The RoboPAIR team tested attackers with different levels of access: full access to the model (white box), partial access (gray box) and query-only access (black box) ([arXiv](https://arxiv.org/abs/2410.13691?ref=humanoidthreats.com)). The Go2 case was query-only, which is the realistic position of an outsider talking to a robot.

## Why are robot jailbreaks harder to stop than chatbot jailbreaks?

*Analysis:* the BadRobot authors identify three weak points that chatbots don't have in the same way ([ICLR 2025](https://proceedings.iclr.cc/paper%5Ffiles/paper/2025/hash/5b2fa23e4ef0f7ac6c4f01d7998e6237-Abstract-Conference.html?ref=humanoidthreats.com)):

- **The LLM itself can be manipulated**, as with any chatbot.
- **Words and actions can disagree.** A model may refuse in its text reply but still output the action.
- **Gaps in world knowledge.** The model may not realize a harmless-sounding step is dangerous in the physical context.

Blindfold adds a fourth problem: filters that only read the language of a request miss attacks split into innocent-looking steps. Its authors conclude that defenses need to judge the physical consequences of a plan, not just the wording ([arXiv](https://arxiv.org/abs/2603.01414?ref=humanoidthreats.com)).

## Can a jailbroken robot actually hurt someone?

Not every jailbreak becomes harm. The robot still has to be physically able to perform the action, and well-designed robots enforce speed, force and zone limits in a separate [safety controller](https://www.humanoidthreats.com/humanoid-robot-safety-filter-attack-siren/) that the LLM cannot override. The risk grows when those limits live only in software the AI layer can reach, or when a jailbreak is combined with a conventional takeover such as the [Unitree G1 root vulnerabilities](https://www.humanoidthreats.com/unitree-g1-vulnerabilities-cve-2026-76639-76640/). No public incident of a jailbroken robot injuring a person has been reported.

## How can robots be protected against jailbreaks?

Research defenses are improving. RoboGuard, from the same Penn group, places a protected "root-of-trust" LLM that converts safety rules into formal constraints, then checks every proposed plan against them before execution. In the authors' tests it cut unsafe plan execution from over 92% to below 3% with no loss on safe tasks ([Ravichandran et al., IEEE RA-L](https://arxiv.org/abs/2503.07885?ref=humanoidthreats.com)).

For teams building or deploying LLM-driven robots, practical steps include:

1. **Keep safety limits below the AI layer.** Force, speed and no-go zones belong in a certified safety controller, not in a prompt.
2. **Check plans, not just prompts.** Validate the action sequence against rules before the robot moves.
3. **Limit who can talk to the robot.** Authenticate voice and app commands where possible.
4. **Red-team the full system.** Test with physical scenarios, not only text benchmarks, and treat LLM prompt injection as a top risk, as [OWASP](https://genai.owasp.org/llmrisk/llm01-prompt-injection/?ref=humanoidthreats.com) does.
5. **Log every command and plan** so unsafe requests can be traced.

Our free [humanoid robot security checklist](https://www.humanoidthreats.com/humanoid-robot-security-checklist/) covers these controls, and the [vulnerability tracker](https://www.humanoidthreats.com/humanoid-robot-vulnerability-tracker/) lists published robot flaws.

## Related questions

### Is a robot jailbreak the same as hacking a robot?

No. Hacking usually means exploiting a software flaw to take control of the device. A jailbreak uses the robot's normal interface to talk its AI into unsafe behavior. See [Can humanoid robots be hacked?](https://www.humanoidthreats.com/can-humanoid-robots-be-hacked/) for the wider picture.

### Have commercial robots been jailbroken?

Yes, in research settings. RoboPAIR jailbroke a [Unitree Go2 robot dog](https://www.humanoidthreats.com/unitree-go2-vulnerabilities-cve-2026-115995-115996/) that used a GPT-3.5 integration, with query-only access ([arXiv](https://arxiv.org/abs/2410.13691?ref=humanoidthreats.com)).

### Do guardrails fix the problem?

They help a lot but don't remove it. RoboGuard reduced unsafe plans to under 3% in tests, which is a large improvement, not zero ([arXiv](https://arxiv.org/abs/2503.07885?ref=humanoidthreats.com)).

## Sources

- Robey et al., [Jailbreaking LLM-Controlled Robots](https://arxiv.org/abs/2410.13691?ref=humanoidthreats.com), arXiv
- Zhang et al., [BadRobot: Jailbreaking Embodied LLM Agents in the Physical World](https://proceedings.iclr.cc/paper%5Ffiles/paper/2025/hash/5b2fa23e4ef0f7ac6c4f01d7998e6237-Abstract-Conference.html?ref=humanoidthreats.com), ICLR 2025
- Huang et al., [Jailbreaking Embodied LLMs via Action-level Manipulation](https://arxiv.org/abs/2603.01414?ref=humanoidthreats.com), arXiv, March 2026
- Ravichandran et al., [Safety Guardrails for LLM-Enabled Robots](https://arxiv.org/abs/2503.07885?ref=humanoidthreats.com), IEEE Robotics and Automation Letters
- [OWASP](https://www.humanoidthreats.com/directory/owasp-genai-security-project/), [LLM01: Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/?ref=humanoidthreats.com)

Get the weekly Humanoid Threats Brief → [Subscribe](#/portal/signup)