When a transform feedback session is (re)prepared in WebGLTransformFeedback::PrepareTransformFeedback, the code recomputes mActive_VertCapacity from the currently bound buffers (minVertCapacity) but did not correspondingly re-clamp the stale mActive_VertPosition carried over from a prior state. This violates the invariant that mActive_VertPosition <= mActive_VertCapacity, which downstream draw/capacity accounting relies on to bound writes into the feedback buffers. If a newly computed capacity was smaller than the leftover position value, later captured vertices would be written starting past the intended buffer boundary, producing an out-of-bounds condition.

The fix adds mActive_VertPosition = std::min(mActive_VertPosition, mActive_VertCapacity);, forcing the position back within the freshly computed capacity so the invariant holds before any vertices are captured. Because the clamp runs on every preparation of the session, any subsequent capacity accounting starts from a valid, in-bounds position.

An attacker running WebGL from web content can drive captured-vertex writes out of the intended buffer bounds within the content/GPU process, corrupting adjacent memory. This requires the attacker to script a transform feedback session and re-prepare it so the recomputed capacity falls below the already-advanced position. The metadata classifies this as a high-impact privilege escalation in the Graphics: CanvasWebGL component.