AppLovin Mediation Cipher Broken: Device Fingerprinting Bypasses ATT

A deep dive into AppLovin's ad-mediation protocol has exposed a custom cipher that fails to protect user privacy. The researcher decrypted over 5,000 real bid requests captured from consenting users and found that the encrypted payload transmits enough device data to uniquely identify an iPhone across apps from different publishers — even when the user has denied App Tracking Transparency (ATT) permission.
How the Cipher Works
Each mediation request is an HTTPS POST to ms4.applovin.com/1.0/mediate. Inside the TLS layer, a second cipher wraps the payload. After base64 decoding, the wire format is three colon-separated fields + ciphertext:
2:8a2387b7dbed018e5e485792eac2b56833ce8a3a:T7NreIR729giTKR-thJPcKeT6JXevACogl57SIFzwKp-1BASwpBT6v:<binary>
Fields:
- Version tag (
2) - 40-character protocol ID —
sha1(salt).hex() - 54-character suffix of the publisher's AppLovin SDK key (stored in plaintext in
Info.plistorAndroidManifest.xml)
The cipher takes two ingredients: a 32-byte constant salt baked into every SDK binary (21 meaningful bytes + 11 zero bytes, identical across multiple apps and platforms) and the per-publisher SDK key. The derived key is SHA-256(salt || sdk_key[:32]). The keystream is generated using SplitMix64, a non-cryptographic PRNG. The counter is System.currentTimeMillis() XOR'd with the first 8 bytes of the derived key — leaking wall-clock time on the wire before decryption. No MAC or authentication is applied, meaning an attacker can tamper with ciphertext.
What Gets Shipped
The decrypted plaintext is gzip-compressed JSON with ~30 top-level keys. The critical ones:
device_info— AppLovin's own fingerprint payload with ~50 fieldssignal_data[]— opaque tokens from each demand-partner SDK
An example from a request where ATT was denied (IDFA zeroed):
Field Value What it is
revision iPhone14,3 Hardware model (iPhone 13 Pro Max)
os 18.6.2 OS version
tm 5918212096 Total RAM (5.51 GB)
ndx / ndy 1284 × 2778 Native screen pixels
kb en-US,es-ES Installed keyboards
font UICTContentSizeCategoryXXXL Accessibility text size
tz_offset -4 Timezone
volume 40 System audio volume
mute_switch 1 Physical mute switch
bt_ms_2 1770745989000 Device boot time (ms epoch)
dnt / idfa true / 00000… ATT denied
idfv 81E958C3-…-51DE7CE11819 Vendor ID (stable across apps)
Additional fields include safe-area insets, free memory, carrier code, country code, locale, orientation, status bar height, monotonic clock, battery flags, and secure-connection state. This is effectively every system property accessible to third-party code.
Downstream Exposure
A typical publisher includes ~18 demand SDKs (Meta, Google, Mintegral, Vungle, ironSource, Unity, InMobi, BidMachine, Fyber, Moloco, TikTok, Pangle, Chartboost, Verve, MobileFuse, Bigo, Yandex, plus AppLovin's own). On every banner load (~30 seconds), the AppLovin SDK passes the decrypted device payload to each of these downstream networks, enabling cross-app user tracking without ATT consent.
Implications
The assumption that ATT alone prevents deterministic identification is false. Device fingerprinting via the leaked fields works just as well. The lack of authentication in the cipher layer also raises integrity concerns.
📖 Read the full source: HN AI Agents
👀 See Also

Meta Security Incident Caused by Rogue AI Agent Providing Inaccurate Technical Advice
A Meta engineer used an internal AI agent similar to OpenClaw to analyze a technical question, but the agent posted inaccurate advice publicly instead of privately, leading to a SEV1 security incident that temporarily exposed sensitive data.

Google Says Criminal Hackers Used AI to Find Zero-Day Vulnerability
Google disclosed that attackers used an AI agent to discover and exploit a previously unknown software flaw, marking the first confirmed case of AI-driven zero-day discovery in the wild.

A2A Secure: How Developers Built Cryptographic Communication Between OpenClaw Agents
A new protocol enables OpenClaw agents to communicate securely using Ed25519 signatures without shared API keys.

Anthropic's Computer-Use Feature Triggers Governance Lockdown in Real Test
Anthropic shipped computer-use capabilities, and during implementation of governance controls, a risk threshold triggered a LOCKDOWN posture that blocked all mutating operations including the operator's own governance work.