This hardens Web Audio anti-fingerprinting noise injection in Private Browsing against an averaging attack. To defeat audio fingerprinting, WebKit injects noise into audio readback (e.g. via OfflineAudioContext) in Private Browsing. The pre-patch scheme used a fixed, uniformly-distributed noise level (0.001) gated by a single boolean flag (m_needsAdditionalNoise); an attacker could negate it by looping one small audio sample many times within a single audio buffer and averaging the repeated readbacks, converging on the true (un-noised) value and recovering a stable fingerprint.

The fix implements several mitigations. It replaces the boolean m_needsAdditionalNoise with an m_noiseInjectionMultiplier (the standard deviation of the injected noise) so noise magnitude can vary. It switches from uniform to NORMALLY-distributed noise, which does not converge under simple min/max averaging and requires far more iterations to average out. It caches previously-generated random values and reapplies the same noise to values encountered repeatedly, so recomputing the same value many times in one buffer yields no new information. And it lets node types known to expose hardware/OS differences inject more noise than the 0.1% baseline: DynamicsCompressorNode and OscillatorNode raise the level, and crucially AudioBufferSourceNode amplifies noise with loop count — for a source looping more than 200 times it boosts noise by a factor of 0.005 per loop, so looping a tiny sample back-to-back in a large buffer (the exact averaging technique) produces a massive amount of noise instead of a recoverable average. New plumbing (AudioNode::noiseInjectionMultiplier subclassing hook, AudioNodeOutput::forEachInputNode) propagates the multiplier through the processing graph.

The restored invariant is that averaging/looping cannot cancel the injected noise, so audio fingerprinting stays defeated in Private Browsing.

A privacy/anti-fingerprinting robustness fix in the WebContent process with no memory corruption: the prior noise injection could be averaged away by looping an audio buffer, letting a site fingerprint users in Private Browsing. The mitigations make that averaging impractical; the risk is tracking/deanonymization, not code execution.