The Basics
What Sightline actually does
Sightline uses your phone's Bluetooth radio to passively listen for Bluetooth Low Energy (BLE) advertisements — small packets that BLE devices broadcast continuously to announce their presence.
Camera-equipped smart glasses (Meta Ray-Ban, Snap Spectacles, and others) broadcast BLE advertisements just like any other Bluetooth device. Sightline compares what it hears against a database of known manufacturer identifiers, service UUIDs, and advertisement payloads specific to camera glasses. When a match is found, it records the sighting. When the same device appears near you repeatedly across different days and locations, Sightline alerts you — because that pattern is consistent with deliberate, sustained proximity.
Everything runs entirely on your device. No audio, no video, no images. Sightline never accesses your camera or microphone. Detection is based purely on Bluetooth radio signals.
Limitations
What Sightline cannot do
It only detects devices in the signature database
Sightline can only detect devices it has been trained to recognize. A camera glasses model released after the last signature update, or one with a sufficiently generic BLE advertisement, will not be detected. The database is updated regularly, but it is never exhaustive.
Devices with Bluetooth off are invisible
If a device has its Bluetooth radio disabled, it broadcasts nothing and cannot be detected. Some glasses allow users to disable BLE independently of other features.
MAC address rotation causes tracking gaps
Modern BLE devices rotate their hardware address (MAC address) every 7–15 minutes to protect user privacy. This is good design — but it means the same physical device may appear to Sightline as many different devices over time. We use content fingerprinting to correlate rotations, but this is imperfect, especially for devices with minimal advertisement payload.
Range and environment affect results
BLE signals travel roughly 10–30 metres in open space, but walls, crowds, and radio interference can reduce this significantly or create ghost readings at the edge of detection range. A single sighting in a crowded environment carries less weight than repeated sightings in varied locations.
It cannot determine if recording is active
BLE advertisements are broadcast regardless of whether the camera is in use. There is currently no reliable, non-invasive way to determine from the outside whether a camera device is actively recording.
False positives are possible
A friend, colleague, or stranger who happens to pass through your environment multiple times wearing camera glasses may trigger a pattern alert even if their presence is entirely coincidental. Sightline reports patterns — it does not interpret them.
Current Approach
How we're handling these limitations today
Rather than hide the gaps, we've built the detection system around a tiered confidence model that acknowledges uncertainty at every step.
Signature matching is the primary signal. When a device's BLE advertisement matches a known camera glasses signature exactly, confidence is high. These matches drive the core alert system.
Content fingerprinting is used to correlate the same physical device across MAC address rotations. When a device's manufacturer payload, service UUIDs, and advertisement structure stay consistent across an address change, we treat them as the same device — increasing detection continuity without relying on the hardware address.
Temporal and spatial pattern analysis is the final layer. A single sighting generates a low-confidence alert. Repeated sightings across different days and independently-derived locations escalate that confidence. This reduces the noise from incidental encounters while surfacing patterns that warrant attention.
The app is explicit about what it knows and doesn't know. Alert language is deliberately cautious — "a known camera-glasses device was detected nearby" rather than "you are being recorded."
Ongoing Research
What we're working on to improve accuracy
The signature database and rule-based detection engine are a strong foundation, but they have a hard ceiling. The next phase of Sightline's detection capability is machine learning inference running entirely on-device.
-
Active
Labeled BLE dataset collection We are building a ground-truth dataset of BLE advertisement vectors from real camera glasses devices, collected during controlled usage sessions (glasses being worn, camera active, device in standby). This dataset will train and validate future ML models. Collection is opt-in and privacy-preserving — no location data, no user identifiers.
-
Planned
BLE advertisement classifier A lightweight on-device model (~50 KB) trained to classify BLE advertisement vectors as camera glasses, non-camera wearable, or ambient device — without requiring an exact signature match. This would allow detection of new or previously unknown camera glasses models based on learned advertisement patterns. The on-device inference infrastructure is built; model training begins once sufficient labeled BLE data has been collected.
-
Planned
RPA rotation correlation improvement Current fingerprinting works well for devices with rich advertisement payloads. We are researching timing-based and statistical methods to improve correlation for devices with minimal payloads, reducing the gaps caused by MAC address rotation.
-
Planned
VR headset disambiguation Some VR headsets (Quest, etc.) share BLE advertisement characteristics with camera glasses. A dedicated classifier trained to separate these categories would reduce false positives in environments where both types of devices are present.
-
Planned
Stalking pattern detection A temporal anomaly model that distinguishes incidental repeated proximity (same building, shared commute) from unusual patterns (multiple independent locations, off-hours) — providing a more nuanced risk signal than simple sighting count.
-
Planned
Optical lens detection (camera required) An optional, opt-in camera-based model that uses the phone's camera to detect lens reflections and IR LED signatures characteristic of recording devices — a second detection vector independent of Bluetooth entirely. This feature will require explicit user consent and will never run without it.
All ML inference will run entirely on-device. No images, no audio, and no BLE data will leave your phone as part of these features. The only data we collect for research purposes is anonymized, opt-in BLE advertisement vectors with labels — never location, never audio, never camera frames.
Questions
Get in touch
If you have questions about how detection works, want to report a device we should add to the signature database, or have feedback on a detection result, reach us at mail@applabs.llc.