This week: researchers turn a humanoid's own safety filter into a collision tool, a trigger-word backdoor for robot task planners, and the FCC grants Intel the first waiver for robots already on a factory floor.
Top threat: when the safety layer is the attack surface
A team from Arizona State University and Carnegie Mellon University showed that the optimization-based safety filters meant to keep humanoids from colliding with people and objects can be pushed into failure using only normal task requests. Their tool, SIREN, searches for goals that drive the robot into states the filter cannot recover from. The paper reports successful attacks against all seven filters tested and collisions on a real Unitree G1 (arXiv 2610.31923).
The filters were research implementations from the open-source SPARK toolkit, not Unitree's shipping firmware, and no CVE was assigned. The root problem is that a filter's safety guarantees hold only when its parameters fit the task and environment, and there is no general way to set them.
Why it matters (analysis): the attacker needs no break-in, only the ability to tell the robot what to do, and the result looks like an accident. Restrict who can assign tasks, check goals against keep-out zones, and re-validate safety parameters at every site. Full story: Humanoid robot safety filters can be turned into collision tools.
New vulnerabilities
- SIREN goal insertion attacks (research, no CVE): normal task requests made a Unitree G1 running research safety filters collide. arXiv 2610.31923
- MuTRAP (research, no CVE): Binghamton University researchers presented a backdoor at IROS 2026 that makes an LLM robot task planner output a harmful plan when a hidden trigger word appears, with 97-100% attack success across seven open models in their tests. The demo used a toy knife and a toy hand on a mobile arm. The team is also working on a defense, PROTEA. arXiv 2504.17070, Binghamton University
Both are now in the Humanoid Robot Vulnerability Tracker.
Incidents
- Unitree H2-X loses a leg on stage: at the World Physical AI Conference in Hangzhou, Unitree's new H2-X humanoid fell while landing a side flip and its right lower leg detached at the joint. No injuries were reported, and Unitree had not commented as of TechNode's report. This was a mechanical failure, not a security incident. TechNode. Our take on the robot itself: Unitree H2-X security risks.
Policy watch
- FCC grants Intel's robot waiver: following up on last week's top story, the FCC's Office of Engineering and Technology on October 22 granted Intel a one-year waiver (DA-26-1146) to make certain hardware changes to foreign-made automated guided vehicles already working in its factories. Wiley describes it as the first Covered List relief for equipment already in the field. Intel committed that the changes will not alter the vehicles' radio characteristics, add new capabilities or swap a US-made component for a foreign one. Wiley
- Why it matters (analysis): operators of installed foreign-made robots now have a precedent for keeping fleets maintained under the FCC rules. Background: FCC robot ban explained.
Tool of the week: SPARK
SPARK (Safe Protective and Assistive Robot Kit) is an open-source toolkit and benchmark from Carnegie Mellon's Robotics Institute for testing humanoid safety controllers in simulation and on a real Unitree G1. It is the testbed the SIREN researchers used, which makes it a practical starting point for research and engineering teams that want to see how their own safety filter settings behave before a robot meets real people. GitHub, paper
Directory pick: HiddenLayer
Backdoors like MuTRAP live in the models that plan a robot's actions. HiddenLayer builds security for AI models and agents across their lifecycle, the layer where tampered models and prompt injection turn into physical risk. Read our HiddenLayer profile or browse the full security directory.
Know someone who should read this? Forward it, or send them humanoidthreats.com/newsletter/
Get the free Humanoid Robot Security Checklist.
Get the weekly Humanoid Threats Brief → Subscribe


