Why the phone is the hard target
When someone is surveilled today, the device that gives them away is usually the phone. It holds the messages, the location history, the microphone and the camera, and it travels everywhere the person does. That is why the serious spyware of the last decade, from mercenary implants sold to governments down to the consumer "monitoring" apps a controlling partner installs, has moved onto Android and iOS.
Finding it is a specific discipline. You cannot trust the phone to report on itself, because the spyware runs with enough privilege to hide. So you work from forensic copies of the device, and you match what you find against known indicators, the way a human rights lab does when a journalist hands over a phone they are worried about.
This is a walk-through of that discipline, end to end. First the toolkit, built from the same open sources those labs publish. Then the method for checking a device you do not trust. Then a teardown of a real surveillance implant, so the indicators stop being abstract.
The toolkit, and where the indicators come from
Three open components do the work.
MVT, the Mobile Verification Toolkit. This is Amnesty International's Security Lab tool, the one used in the original Pegasus Project forensics. It reads a forensic copy of a phone, an iOS backup, an Android bug report, an androidqf acquisition, and checks every artifact against indicators of compromise. Installed here as version 2026.11.2.
The indicator feeds. MVT ships knowing where to pull indicators from, and one command downloads them. On this build it retrieved 18 separate indicator sets covering the mercenary-spyware ecosystem: NSO Group's Pegasus, Intellexa's Predator, Candiru, RCS Lab, QuaDream's KingSpawn, the NoviSpy campaign in Serbia, Operation Triangulation, and more. Deduplicated, that is 4,980 network domains, 73 IP addresses, 83 malicious process names and 65 email addresses, each one tied to a documented spyware operation.
The stalkerware set. Mercenary spyware is only half the threat. The other half is commercial "stalkerware", apps sold openly to monitor a partner or a child, which behave like spyware once installed. The Echap stalkerware-indicators project tracks them, and it is substantial: 133 families, 1,659 package identifiers, 473 signing-certificate fingerprints, 929 command-and-control hosts, 3,228 sample hashes, and 12 YARA rules. Families like FlexiSpy, mSpy, Cerberus, TheTruthSpy and Hoverwatch are all in it.
Between the two, you are matching a suspect device against roughly five thousand mercenary-spyware network indicators and the full fingerprint of 133 consumer-stalkerware families. That is the corpus. The skill is in using it without fooling yourself.
Checking a device you do not trust
The rule in device forensics is simple: never ask the phone a question the spyware can lie about. You do not open the suspected app and read its settings. You take a forensic copy off the device and analyse the copy.
On Android, MVT works from four kinds of acquisition, in rough order of how much they reveal:
- A full bug report (
check-bugreport), which the phone can generate itself and which captures running services, recent network use and battery history. - An `androidqf` acquisition (
check-androidqf), a structured dump of installed packages, their certificates, running processes and system logs. - A backup (
check-backup), which can include SMS databases where some spyware leaves its command messages. - A live ADB session (
check-adb), used only on a device the owner has unlocked and consented to, which lists every installed package and lets you pull the ones that look wrong.
Each one gets checked against the full indicator corpus above. A hit is not a verdict on its own, it is a lead: a known Pegasus domain in the network history, a package whose signing certificate matches a stalkerware vendor, a process name that should not exist. The analyst's job is to corroborate the lead across more than one artifact before calling it.
This is consent-first work. The method assumes the device belongs to someone who asked for help, which is exactly the civil-society threat model: a journalist, an activist, a defender who thinks their phone is watching them. It is the opposite of covert.
Taking a real implant apart
To make the indicators concrete, here is a static teardown of a real surveillance sample: Skygofree, an Android implant first documented by Kaspersky in 2018 and linked to an Italian vendor. The sample came from a public malware research archive. Everything below was read from the file statically, by parsing it. The sample was never installed or run.
The disguise comes first. The app declares itself as "System Update", package name com.sysmanager, version 3.4.3. A victim who glances at the app list sees a system component, not an app. (A second sample in the same family poses as "Syncsystem", core.syncsystem.)
The permission request tells you what it is built to take. Parsed from the manifest, 24 permissions, and the surveillance-relevant ones group cleanly:
- Microphone and camera:
RECORD_AUDIO,CAMERA - Location:
ACCESS_FINE_LOCATION,ACCESS_COARSE_LOCATION - Messages and calls:
READ_SMS,RECEIVE_SMS,READ_CALL_LOG,PROCESS_OUTGOING_CALLS,READ_PHONE_STATE - Contacts:
READ_CONTACTS,GET_ACCOUNTS - Persistence:
RECEIVE_BOOT_COMPLETED,WAKE_LOCK
That is the whole surveillance surface of a phone in one manifest: what you say, where you are, who you talk to and what you write.
The component list shows how it is organised. The implant registers 25 background services and 10 broadcast receivers. Reading the class names is like reading a feature list:
AndroidCamerafor photo and video captureAndroidAccessibility,AccessibilityServiceandNotificationListener, the accessibility-abuse technique that lets an implant read what other apps show on screen. This is how Skygofree reaches into messaging apps without breaking their encryption: it reads the screen after the message is decrypted for the user.CallReceiverandAndroidCallSystemfor call handling,SmsReceiverfor text interceptionAndroidDeviceAdministrator, which claims device-admin rights so the app is harder to uninstallFetchAddressIntentServicealongside geofencing code, Skygofree's signature trick: start recording audio automatically when the victim enters a chosen locationAndroidMDMSupport$Social, code aimed specifically at social and messaging apps
The command channel is in there too. The sample carries a hardcoded registration endpoint, http://url.plus/app/pro/register.php, and a Firebase project, future-8a57f.firebaseio.com, used to push commands over Google's own messaging service so the traffic blends in. Both are concrete network indicators you could now hunt for across a fleet.
Finally, attribution. The signing certificate is not anonymous: its organisational unit is literally negg, the vendor name researchers tied to Skygofree, with a SHA-1 of 73EDF6B453CE760EBC9AF20D3FBD23F326F5DC9E and validity dates stretching to 2042. A matching certificate on another sample is strong evidence it came from the same workshop. That single fingerprint is a better pivot than any filename.
The result that matters most: the honest negative
Run the consumer-stalkerware checks against this sample and they come back clean. The 12 YARA rules do not fire. The package name and certificate are not in the 1,659-package, 473-certificate stalkerware corpus.
That is not a failure of the tooling. It is the point. Skygofree is a targeted mercenary implant, not a commercial stalkerware product, so it lives in MVT's mercenary-spyware indicator feeds, not in a consumer blocklist. An analyst who only ran the stalkerware rules would have declared a surveillance implant clean.
The lesson is the one that separates real forensics from box-ticking: a clean result is only as good as the indicator set behind it, and no single set is complete. You match against mercenary feeds and stalkerware feeds and the device's own behaviour, and you treat "no hit" as "no hit from this corpus", never as "safe".
For comparison, the same pipeline run against a known-good app returns exactly what it should: a two-permission app, no background services, no surveillance capability, no corpus match. A clean baseline is what lets a dirty result stand out.
What this is, in one line
The threat is a phone that watches its owner. The defence is forensic copies, a maintained indicator corpus, and an analyst who corroborates before concluding and never trusts a single clean scan. The tools are open. The discipline is the work.