The Royal Institution's She tested 210 smart devices. Here's what they're hiding | IoT with Anna Maria Mandalari: skim's analysis identifies 10 key moments. UCL lecturer Anna Maria Mandalari details laboratory experiments on over 210 Internet of Things devices, revealing widespread data collection, misactivations, and security flaws.
Category: Tech. Format: Educational. YouTube video analyzed by skim.
skim AI Analysis
Credibility assessment: Empirical Lab Research. Anna Maria Mandalari presents empirical findings from years of packet-level testing across more than 210 connected devices in two university testbeds at UCL and Northeastern University, detailing reproducible methodologies and specific vendor behaviors.
Bias assessment: Consumer Privacy Advocate. The lecture strongly prioritizes user privacy and security over commercial data collection, framing manufacturer tracking and cloud processing as adversarial while advocating for local edge-based technical interventions.
Originality: 86% — Primary Experimental Data. Rather than summarizing third-party news, the speaker shares original laboratory experiments covering smart speakers, TV automated content recognition, AI browser extensions, Bluetooth medical sensors, and commercial IoT safeguards.
Depth: 85% — Systematic Packet Analysis. The talk explains specific technical mechanisms behind device surveillance and vulnerability, including network traffic patterns, Bluetooth packet bit manipulation, ARP spoofing overhead, and edge AI traffic blocking.
Key Points (10)
1. Mandalari reveals UK IoT data flows abroad
Timestamp: 00:01:36 to 00:04:26
Testing over 210 connected devices across labs at University College London and Northeastern University since 2018 shows that most traffic from UK devices routes to servers in the US or China outside European privacy jurisdictions. Even when users opt out of motion recording on smart doorbells or lack accounts on LG and Samsung TVs, the devices still record and contact third parties like Netflix, Facebook, and Google.
Significance (High): Cross-border routing exposes European households to foreign surveillance regimes and weakens local privacy protections when opt-out settings fail to halt data transmission.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
2. Smart speakers record up to 20 seconds on errors
Timestamp: 00:06:18 to 00:09:11
By streaming over 500 hours of Netflix audio to smart speakers in an automated test cabinet, researchers found that devices including Microsoft Invoke, Echo 2, and Apple HomePod frequently misactivate on phrases sounding like wake words. During these unintended triggers, speakers recorded and uploaded up to 20 seconds of audio to cloud servers without the user realizing.
Significance (High): Frequent false wake-word triggers turn local voice assistants into intermittent cloud microphones capable of capturing full household conversations.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
3. Smart TVs snapshot screens even after opt-out
Timestamp: 00:12:01 to 00:14:56
Smart TVs bought in the last 10 years enable Automated Content Recognition by default to fingerprint viewing habits, with Samsung capturing screen snapshots every 500 milliseconds and LG capturing audio every 10 milliseconds. Automated tests in the UK and US revealed that ACR continues running and sending fingerprints to vendor servers and third parties like Google Analytics even when users opt out or use the TV solely as an external monitor for a laptop or PlayStation.
Significance (High): Using a smart TV as a computer monitor exposes personal laptop screens to sub-second vendor fingerprinting despite explicit user opt-outs.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
4. AI browser extensions harvest medical and ID data
Timestamp: 00:17:47 to 00:21:36
Across more than 1,000 automated persona tests on generative AI browser extensions such as Sider AI, Monica, Merlin, MaxAI, and Perplexity, Mandalari's team found that the tools capture full webpage DOMs, screenshots, and web form inputs. These extensions collected sensitive attributes including national security numbers, NHS medical health records, and UCL student records, sometimes sharing full chat transcripts with third parties like Google Analytics.
Significance (High): Granting browser extensions page-reading permissions exposes confidential health and identity portals to cloud scraping and cross-account profiling.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
5. Mandalari hacks Bluetooth medical sensors for 15 pounds
Timestamp: 00:21:48 to 00:26:45
Testing wearable medical devices such as smart oximeters, blood pressure monitors, and Bluetooth Low Energy glucose sensors linked to insulin pumps revealed that every tested device was vulnerable to at least one attack using a 15-pound Amazon dongle. From up to 30 meters through walls or 100 meters in open space, researchers intercepted packets, performed man-in-the-middle spoofing to alter displayed blood oxygen from 99% to 67%, and disabled glucose sensors via denial-of-service attacks.
Significance (High): Unauthenticated Bluetooth Low Energy traffic in closed-loop medical wearables turns cheap off-the-shelf radio dongles into life-threatening weapons.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
6. IoT security boxes leak data and misidentify gear
Timestamp: 00:28:30 to 00:31:52
To evaluate commercial IoT security safeguards—including standalone boxes, router software, bridge devices, and ARP-spoofing gadgets—the lab connected all 210 IoT devices and emulated network attacks. Instead of purely protecting privacy, safeguards like Avira and Fingbox contacted third-party tracking services such as Mixpanel without disclosing them in their privacy policies.
Significance (Medium): Consumers buying smart-home security appliances often introduce new third-party trackers into their routers without gaining accurate device visibility.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
7. Commercial safeguards miss most network attacks
Timestamp: 00:31:53 to 00:37:39
Across 30 automated repetitions per attack type with 20-minute detection windows over a full year, commercial IoT safeguards failed to detect the vast majority of simulated threats, including basic port scanning. Even when Bitdefender detected an HTTP flooding denial-of-service attack, the alert arrived after three minutes, which can be too late to prevent disruption.
Significance (High): Relying on commercial router safeguards creates a false sense of security while degrading network speed and shipping home metadata to cloud servers.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
8. Mandalari warns voice AI enables emotional profiling
Timestamp: 00:37:39 to 00:40:03
Mandalari argues that routing UK device data to support and third-party servers matters because aggregated metadata enables deep behavioral profiling and mass manipulation similar to the Cambridge Analytica scandal. With artificial intelligence, voice assistants can infer user emotions, stress, and depression levels from vocal tone, an area where Amazon already holds a patent.
Significance (Medium): Combining continuous smart-home sensing with AI voice analysis shifts surveillance from targeted advertising toward predictive emotional and behavioral manipulation.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
9. Edge AI router blocker cuts non-essential traffic
Timestamp: 00:40:04 to 00:43:33
To protect privacy without relying on cloud servers, Mandalari's lab developed an edge AI system installed directly on the home router to test whether network destinations are essential for device functionality. Testing revealed that at least half of the complex IoT devices—such as cameras, smart speakers, and TVs—contact non-essential domains, most of which belong to third parties, and continue working normally when those domains are blocked.
Significance (High): Running local traffic classification on home routers proves that consumers can strip out third-party IoT tracking without breaking remote device features.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
10. Smart plugs detect cyberattacks via power spikes
Timestamp: 00:43:35 to 00:48:15
By connecting all lab IoT devices to energy-monitoring smart plugs sampling power consumption once per second, researchers trained a machine learning model to spot ongoing cyberattacks. For the majority of devices—excluding high-variance streaming hardware like smart TVs—denial-of-service, man-in-the-middle, and scanning attacks produced recognizable peaks and patterns in electrical draw.
Significance (Medium): Monitoring electrical draw at the plug offers a privacy-preserving way to detect active device compromises without inspecting encrypted network payloads.
Sources in support: Anna Maria Mandalari (Lecturer at UCL and Director of the SafeNetIoT Laboratory)
This analysis was generated by skim (skim.plus), an AI-powered content analysis platform by Credible AI. Scores and classifications represent the platform's AI-generated assessment and should be considered alongside other sources.