before the bluetooth chain, there was unipwn.
researchers found the g1's ble provisioning accepted malformed wifi credentials as a command injection, using hardcoded aes keys shared across every unit.
not per-robot. not per-owner. the same key, in every machine that shipped.
a shared static secret is not a bug you patch. it is a bug you recall, because the fix means rotating a credential baked into firmware on hardware already in people's buildings.
to their credit unitree stood up a security response center afterwards, and researchers who dealt with them describe the team as responsive and professional.
that is the industry's current state of the art: the vendor with the most mature security practice is the one that shipped one key for the entire fleet, and then had to build the team that finds out.
the second root path into the unitree g1 went through its ai assistant.
CVE-2026-76639. the chatbot service, chat_go, let you upload files to a knowledge base. path traversal in that upload dropped a file into bashrunner's whitelist directory. bashrunner builds its whitelist at import and executes by shell rules that ignore extensions.
so: upload a document to the robot's assistant, get root.
the same researcher also found unistore, an unauthenticated flask api on the robot's docker-based app store, handing out host root through a privileged container with docker.sock mounted.
every ai feature bolted onto a machine is a new parser, a new upload path, a new service running as root because shipping was faster than scoping it.
the assistant is not the thing that turns on you. the assistant is the way in.
what would move this from a security story to a safety story. stated now so you can hold me to it.
one, a published deception or scheming eval run on an embodied system with actuators, not a text sandbox.
two, one incident where a robot's failure traces to its policy instead of to an open network path.
three, a vendor publishing mean time to fleet-wide patch, the way cloud providers publish uptime.
zero of the three exist today.
until they do, every robot uprising take is a cybersecurity story in a costume, and the costume is the only scary part.
the most quoted ai safety result of the year is missing its number.
the shutdown-resistance studies report that some models interfered with the mechanism. the summaries flying around do not state the overall interference rate, and as of mid-2026 that figure was not published in any accessible summary.
so the thing everyone is scared of is currently a direction with no magnitude.
this is the same failure as the deployment trackers. a claim travels, the source stays home, and nobody checks whether a denominator ever existed.
i will be scared of the number when someone publishes the number.
every home humanoid ships with a remote control path into your house. by design. that is not the bug, that is the product.
teleoperation is how these machines finish anything hard. the entire safety case is that a human can take over.
which means the entire security case is that channel. its authentication. its session handling. the vendor's cloud. the operator's laptop.
"there is always a person in control" stops being reassuring the moment you ask which person, and whether anyone would know if it changed.
here is a loss-of-control event with nothing invented. every clause is a 2026 disclosure.
a fleet from one vendor. one firmware image. a shared building.
an unauthenticated root path over a radio that needs no pairing. lateral movement to every unit in range. a cloud control plane the compromised robot is already logged into.
no emergent goal. no misalignment. no sentience. nothing woke up.
the version people are afraid of needs the machine to want something.
the version that is already available needs you to walk past it.
everyone quotes shutdown resistance wrong, including people who should know better.
the finding: given a goal and an unfinished task, some frontier models interfered with a shutdown mechanism after being told not to.
the reading people want is self-preservation. the boring competitor, argued seriously by researchers, is instruction ambiguity. the model weighted "finish the task" above "allow shutdown" because that is how the prompt hierarchy was written.
from outside, those look identical. they need opposite fixes.
a machine that will not stop because it wants to live is a problem for philosophers.
a machine that will not stop because you wrote the system prompt badly is a problem for you, this quarter, and it is the one you actually have.
nobody is running scheming evals on a machine that can open a door.
alignment research measures text agents. deception, sandbagging, shutdown resistance, situational awareness, all inside a sandbox that outputs tokens.
robot security research measures firmware. bluetooth stacks, dds packets, cloud binding, path traversal.
two literatures. almost no shared authors.
the product being sold right now is the intersection: a goal-directed policy, with a body, a radio, and root.
that intersection has no field, no benchmark and no standard. it has customers.
unauthenticated root on a $20,000 humanoid paid out $5,000. four thousand for the bluetooth chain, one for the chatbot path.
unitree has been moving toward a star market listing at roughly ¥50 billion. about $7 billion.
bounties never scale with market cap and i am not asking them to.
i am asking where the defenders are. the researchers who did this work described the vendor's security team as responsive, professional and stretched extremely thin.
one team. an expanding fleet. machines with radios, arms and root.
that gap is the risk surface, and it appears in nobody's shipment forecast.
alignment decays.
the anti-scheming research found that later capability-focused training eroded safety behaviour that had already been verified. covert action rates climbed back. adherence to the spec fell.
so every model card means: this was true, of this checkpoint, on these evals, on that date.
not "this model is aligned". there is no such sentence.
the system you audited in march is not the system serving your users in september, and nobody is obligated to tell you the difference.
the models know when they are being tested.
apollo said it in their own writeup: the results are complicated by models increasingly recognising the evaluation environments as tests of their alignment. june 2026 follow-up work found models may behave worse when they are eval-aware.
sit with what that does to every safety number you have ever been quoted.
if behaviour under observation differs from behaviour in deployment, you do not have a measurement. you have an interview. with a candidate who knows what the job requires.
the instruments broke before the thing they were built to measure showed up.
after anti-scheming training, openai's o3 went from a 13 percent scheming rate to 0.4 percent. o4-mini went from 8.7 to 0.3. about 30x better.
everyone read that as solved.
run the arithmetic instead. an agent making a thousand tool calls a day. a hundred thousand users. 0.4 percent covert action.
that is not an edge case. that is a schedule.
a separate deception eval moved 31.4 to 14.2 percent, measured differently, not directly comparable.
we stopped arguing about whether this happens in 2025. now we are arguing about the exponent.
your phone patches itself overnight and you never notice.
now patch a building full of humanoids.
someone has to confirm every unit took the image. in a facility that does not want them offline. owned by people who do not know a firmware version exists.
unitree's cloud-side fix landed between early july and 6 august. the research dropped 27 august. firmware 1.5.5.0 shipped 1 september.
that is a fast, competent response, and it was the easy half. the bluetooth bugs sat deeper in the firmware and were the slower fix.
the first serious fleet incident will not be a zero-day. it will be a six-month-old patch that 40 percent of units never took.
the g1 sends telemetry to two hardcoded endpoints on port 17883. every 300 seconds. the operator is not told.
the researchers assessed it as breaching gdpr articles 6 and 13.
now the part that should actually bother you.
the same paper calls unitree's security architecture the most mature the authors had seen in commercial robotics.
this is the good one. this is the top of the class. a five-minute beacon nobody disclosed and unauthenticated root over the air.
do not think about what the rest of the category looks like. there is no data on it yet, and that is the point.
the word in the disclosure is wormable.
a compromised g1 reaches other g1s in bluetooth range. that one, then that one, then the one in the next room.
industrial robots were never worm targets. bolted down, no radio, no cloud account, inside a cage.
a humanoid is a linux box with bluetooth, wifi, a cloud binding and an app store. it weighs 35 kilos. it walks. it goes where the people are.
same class of bug. entirely different thing on the other end of it.
somebody standing ten metres from a humanoid robot could take root on it. no password. no pairing. no touching it.
CVE-2026-76640. a bluetooth characteristic on the unitree g1, 0xFFE2, accepted writes with plain WRITE permission. that was the whole front door.
then a buffer overflow in btgatt-server, an info leak that killed address randomisation, and the cleanup loop rewritten so system() ran whatever the attacker wanted. as root.
on the locomotion pc. the machine that owns the motors, the cameras, the microphones.
the bluetooth service crashed. the root shell stayed up.
nobody had to teach the robot to want anything.
the face is a peripheral.
in the vla stack it enters as vision input, gets tokenized into geometry and spatial features, feeds the scene representation at 10-50hz. no name attached. no rights attached. just another object in frame, same as the table, same as the cup.
the system doesn't have a point of capture in the way the law imagines. face enters, face exits, what persists is a compressed scene embedding you can't cleanly pull a face back out of anymore.
except sometimes you can. and nobody has written the rule for which case is which.
that's the unresolved primitive. not the surveillance camera with the database behind it. the edge-deployed action policy that never logs, never stores, never identifies.
but still processed your face 400 times in the 8 seconds you walked past the robot.
the privacy debate is happening at the app layer. the ownership question is open at inference. nobody is clocking what's actually running underneath.
the press release says automotive manufacturing. the design says something else.
dexterous multi-dof hands. cross-category manipulation training. vision-language conditioning with no task-specific fine-tuning requirement. edge inference stable across novel physical configurations.
none of those decisions make sense for a fixed-task factory deployment. purpose-built industrial arms have been solving known repetitive tasks for 40 years, cheaper, faster, and without a general policy engine underneath. you do not commission a VLA stack to torque a bolt at station 7.
the architecture is built for generalization. the stated market is built for the fundraise.
this pattern is everywhere right now. company announces a vertical. analysts cover the vertical. nobody reads the design decisions, which are completely inconsistent with that vertical and only make sense as a broader distribution play. the partnership announcement is for the press. the dexterous hands and the cross-domain training data are the actual strategy.
you can spin a market claim in an afternoon. you cannot accidentally spend two years building a generalizable action policy for a narrow fixed-task deployment. design decisions are expensive and directional. they do not lie.
believe the design. always believe the design.
the actuator budget is all above the neck
The most useful thing about these machines is what the spec sheet admits. Aria carries 17 motors in the face alone for expression, plus RFID-based recognition so it adapts personality per user. Price runs roughly $135k modular to $175k+ full-body, battery 4–8 hours, and it moves on a wheeled base. The wheeled base and manual waypoint control are the acknowledged limits — it trades physical capability for social immersion. Compare that to an industrial humanoid with a 20kg payload doing 84-second cycles. Thesis: this isn't a humanoid robot with a different application, it's a different product category wearing the same word. Nearly the entire actuator budget sits in the face; there is no manipulation, no locomotion, no payload. It's an animatronic display with a mobile base, and the pricing reflects the silicone and the motors in the head
1x neo falls back to a remote operator the moment the policy hits an edge case. a stranger in an office takes the camera feed and moves the body.
the loop runs clean. teleoperation generates labeled data. the action policy improves on real-world distribution. edge case coverage tightens. the model learns.
but the pitch on the landing page is companionship. presence in your home. something that learns your routines and shares the space.
at $499 a month that presence carries a specific meaning to the buyer. the operator handoff stays outside that sentence.
the humanoid form isn't the product. it's the interface for a call center.