Sammy Azdoufal wanted to steer his robot vacuum with a PS5 controller. Like any good maker, he thought it would be fun to drive a new DJI Romo around manually. He ended up gaining access to an army of robotic cleaners that gave him eyes into thousands of homes.
Driven by purely playful reasons, Azdoufal used Anthropic’s Claude Code AI coding assistant to reverse-engineer his Romo’s communication protocols. But when his homebrew app connected to DJI’s servers, roughly 7,000 robot vacuums across 24 countries started answering.
He could watch their live camera feeds, listen through onboard microphones, and generate floor plans of homes he’d never visited. With just a 14-digit serial number, he pinpointed a Verge journalist’s robot, confirmed it was cleaning the living room at 80% battery, and produced an accurate map of the house from another country.
The technical failure was almost comically basic. DJI’s MQTT message broker had no topic-level access controls. Once you authenticated with a single device token, you could see traffic from others device in plaintext.
It wasn’t only vacuums that answered back. DJI’s Power portable battery stations, which run on the same MQTT infrastructure, also showed up. These are home-backup generators expandable to 22.5kWh, marketed for keeping your house running during outages.
What makes this different from a conventional security discovery is how it happened. Azdoufal used Claude Code to decompile DJI’s mobile app, understand its protocol, extract his own authentication token, and build a custom client.
AI coding tools are lowering the bar for advanced offensive security. The population capable of probing Internet of Things (IoT) protocols just got much, much larger, further eroding any remaining faith in security through obscurity.
Why plenty of IoT vacuum cleaners suck
This isn’t the first time someone has remotely pwned a robot vacuum cleaner. In 2024, hackers commandeered Ecovacs Deebot X2 vacuums across US cities, shouting slurs through speakers and chasing pets around. Ecovacs’s PIN protection was checked only by the app, never by the server or the device.
Last September, South Korea’s consumer watchdog tested six brands. While Samsung and LG fared well, and found serious flaws in three Chinese models. Dreame’s X50 Ultra allowed remote camera activation. Researcher who Dennis Giese later reported a TLS vulnerability in Dreame’s app to CISA. Dreame didn’t respond to CISA’s queries.
The pattern keeps repeating: manufacturers ship vacuums with textbook security failures, ignore researchers, then scramble when journalists publish.
DJI’s initial response made things worse. Spokesperson Daisy Kong told The Verge the flaw had been fixed the prior week. That statement arrived about thirty minutes before Azdoufal demonstrated thousands of robots, including the journalist’s own review unit, still reporting in live. DJI later issued a fuller statement acknowledging a backend permission validation issue and two patches, on February 8 and 10.
DJI said that TLS encryption was always in place, but Azdoufal says that protects the connection, not what’s inside it. He also told The Verge that additional vulnerabilities remain unpatched, including a PIN bypass on the camera feed.
Regulators are applying pressure
Regulation is arriving, slowly. The EU’s Cyber Resilience Act will require mandatory security-by-design for all connected products sold in the bloc by December 2027, with fines up to €15 million. The UK’s PSTI Act, in force since April 2024, became the world’s first law banning default passwords on smart devices. The US Cyber Trust Mark, by contrast, is voluntary. These frameworks technically apply regardless of where the manufacturer sits. In practice, enforcing fines on a Shenzhen company that ignores CISA coordination requests is a different proposition entirely.
How to stay safe
There are practical steps you can take:
- Check independent security testing before buying connected devices
- Place IoT devices on a separate guest network
- Keep firmware updated
- Disable features you don’t need
And ask yourself whether a vacuum really needs a camera. Many LiDAR-only models navigate effectively without video. If your device includes a camera or microphone, consider whether you’re comfortable with that exposure—or physically cover the lens when not in use.
We don’t just report on threats—we remove them
Cybersecurity risks should never spread beyond a headline. Keep threats off your devices by downloading Malwarebytes today.
