Researchers from Arizona State University and Carnegie Mellon University have shown that a humanoid robot safety filter, the software layer meant to stop collisions, can be steered into failure using nothing more than normal task requests. In a paper posted to arXiv on October 22, 2026, the team describes "goal insertion attacks" and an automated tool called SIREN that finds them, and reports that SIREN-generated attacks made a real Unitree G1 humanoid collide while executing tasks (arXiv 2610.31923).
What happened
The paper, "Not an Accident: Exploiting Safety Filter Inadequacy in Humanoid Robots," targets the optimization-based safety filters that sit between a humanoid's controller and its motors. These filters come with mathematical guarantees against collisions, but the authors point out that the guarantees only hold when the filter's safety parameters are configured correctly for the task and environment, and that there is no general method to find the right settings. A filter tuned at the factory can therefore fail after deployment.
The researchers turned that gap into an attack. Instead of hacking the robot, the attacker submits legitimate task requests through the robot's normal user interface, choosing goals or task locations that drive the robot into a state its safety filter cannot recover from. According to the abstract, the attack needs no extra privileges.
Technical details
- Filters tested: seven optimization-based safety filters, including Control Barrier Functions (CBF), the Safe Set Algorithm (SSA), the Sublevel Safe Set (SSS) and slack-relaxed variants of each. The paper reports successful attacks against every filter it evaluated.
- Testbed: the experiments use 35 vulnerable scenarios from SPARK, an open-source humanoid safety toolkit and benchmark from Carnegie Mellon's Robotics Institute that is built around the Unitree G1 (SPARK paper, code on GitHub).
- SIREN: an automated search tool that finds which goals to insert for a given task and environment. The authors report it found attacks in all vulnerable SPARK scenarios and found substantially more attacks than a baseline search method.
- Real hardware: SIREN-generated attacks caused a physical Unitree G1 to collide during task execution.
The filters in the study are research implementations from SPARK. The paper does not report a flaw in Unitree's own shipping firmware, no CVE has been assigned, and we found no vendor statement on the work. This is published research, not a product vulnerability. We cover it at a defender's level only.
Who is affected
- Teams building on optimization-based safety filters, such as CBF or safe-set methods, whose parameters are fixed for one setting and then reused in others.
- Operators who let many people assign tasks to a humanoid, for example through a fleet dashboard, an app or a voice or chat interface. The attack relies on the ability to request goals, not on network access to the robot's internals.
- Buyers of humanoids that advertise software safety layers as their main protection against collisions with people and objects.
Why it matters for humanoid security
Analysis: Most advice on securing AI-driven robots, including ours, says to keep safety limits outside the AI layer, so a jailbroken LLM or a prompt-injected VLA model cannot override them. That advice still holds. This paper adds a harder lesson: the safety layer itself is an attack surface. If its parameters were never validated for the real environment, an attacker who only controls what the robot is asked to do may be able to make the safety layer fail, and the result looks like an accident.
That matters for liability and incident response too. A collision caused this way leaves no malware and no break-in, only a sequence of ordinary-looking task requests. It belongs on any humanoid robot attack surface map next to sensors, OTA updates and the AI brain.
This month, a separate Binghamton University team publicized MuTRAP, presented at IROS 2026, a backdoor attack that makes an LLM task planner output a harmful plan when a hidden trigger word appears; it was shown on a mobile arm, not a humanoid (arXiv 2504.17070). Together, the two papers show attacks moving up the stack, from firmware to the planning and safety logic that decides how a robot moves.
What to do now
Analysis, based on the research:
- Treat task goals as untrusted input. Check requested goals and waypoints against approved work zones, keep-out areas and people-present rules before they reach the planner.
- Limit who can assign tasks. Require authentication and roles for goal submission, and rate-limit unusual sequences of requests.
- Validate safety parameters per site. Test the safety filter against adversarial goal search in a digital copy of each real environment, and re-test when the layout, payload or task changes. Open benchmarks such as SPARK can help research teams start.
- Keep hardware layers independent. Force, speed and zone limits in a safety-rated controller and a physical E-stop should not depend on the software filter. See our humanoid robot security standards guide for ISO 10218 and ISO 25785-1.
- Log goals, not just faults. Keep a record of every task request alongside collision and near-miss events, so investigators can tell an accident from an attack.
- Ask vendors how their safety parameters were validated and whether they test against adversarial task sequences.
We have added this research to the Humanoid Robot Vulnerability Tracker. For a full deployment review, use the Humanoid Robot Security Checklist.
Sources
- Bao et al., "Not an Accident: Exploiting Safety Filter Inadequacy in Humanoid Robots," arXiv 2610.31923 (Oct 22, 2026)
- SPARK: Safe Protective and Assistive Robot Kit, Carnegie Mellon University (arXiv 2502.03132)
- SPARK source code (GitHub)
- Al Nahian et al., "MuTRAP: Multi-trigger Trojans Attacking Robot Task Planning Systems," arXiv 2504.17070
- Binghamton University: Secret trigger words could make AI-powered robots dangerous
Get the weekly Humanoid Threats Brief → Subscribe


