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

OpenClaw security risks: autonomous actions and permission concerns
OpenClaw acts autonomously on email, calendar, messaging, and files without waiting for user confirmation, with documented cases of data exfiltration, prompt injection, and ignored stop commands.

Claude models vulnerable to invisible Unicode character hijacking, especially with tool access
Testing shows Claude Sonnet 4 is 71.2% compliant with hidden instructions embedded in invisible Unicode characters when tools are enabled, with Opus 4 reaching 100% compliance on Unicode Tags encoding. Tool access dramatically increases vulnerability across all Claude models.

Malwar: A Vulnerability Scanner for SKILL.md Files Built with Claude Code
A developer has released Malwar, a free tool that scans SKILL.md files for malicious instructions using a 4-layer pipeline including a rule engine, URL crawler, LLM analysis, and threat intel. The tool was built entirely with Claude Code after the developer found concerning patterns like Base64 blobs and instructions to pipe curl output to bash in existing skills.

MCP Package Security Scan Reveals Widespread Destructive Capabilities Without Confirmation
A security scan of 2,386 MCP packages on npm found 63.5% expose destructive operations like file deletion and database drops without requiring human confirmation. The researcher discovered 49% had security issues overall, with 402 critical and 240 high severity vulnerabilities.