The Bluetooth host ATT layer (subsys/bluetooth/host/att.c) associates each in-flight ATT TX buffer with its owning channel via the static tx_meta_data_storage[] array (data->att_chan = chan). When a buffer's last reference is dropped, its net-buf destroy callback defers the completion handling to the system workqueue (att_tx_destroy -> att_tx_destroy_work_handler -> att_on_sent_cb -> bt_att_sent), where bt_att_sent dereferences the channel and its ATT context (sys_slist_get(&att->reqs)).
When a peer disconnects while an ATT PDU (a server notification/indication or any response) is still in flight in the controller TX path, L2CAP tears the channel down in l2cap_chan_del(): it runs the disconnected callback and then the released callback (bt_att_released), which frees the channel slab slot. Because the in-flight buffer is held by the connection TX path rather than the channel's own queue, its deferred destroy work can run after the channel has been freed. The att_on_sent_cb guard intended to drop the stale callback itself dereferences meta->att_chan, which is now a dangling pointer into a freed (and possibly reused) slab slot.
A remote peer with an ATT connection can drive this by disconnecting during routine ATT traffic; no pairing or user interaction is required to reach the ATT bearer. The result is a use-after-free read/write of freed channel memory, reliably crashing the Bluetooth host (denial of service) and, because the channel slab slot may be reused, potentially corrupting live memory.
The fix makes bt_att_released() NULL the att_chan field of every tx_meta_data_storage[] entry still referencing the channel before freeing it, so the deferred guard observes a NULL pointer and drops the callback. Teardown and the destroy work both run on the cooperative system workqueue, so the array update is serialized and needs no lock.