diff --git a/crates/unbounded-spsc/RUSTSEC-0000-0000.md b/crates/unbounded-spsc/RUSTSEC-0000-0000.md
new file mode 100644
index 0000000000..7f1063bf4f
--- /dev/null
+++ b/crates/unbounded-spsc/RUSTSEC-0000-0000.md
@@ -0,0 +1,64 @@
+```toml
+[advisory]
+id = "RUSTSEC-0000-0000"
+package = "unbounded-spsc"
+date = "2026-05-14"
+url = "https://github.com/spearman/unbounded-spsc/issues/4"
+categories = ["memory-corruption"]
+keywords = ["transmute", "uaf", "race", "drop"]
+informational = "unsound"
+aliases = ["GHSA-6m57-8r3p-pqx6"]
+
+[versions]
+patched = [">= 0.3.0"]
+unaffected = []
+
+[affected]
+functions = { "unbounded_spsc::Sender::send" = ["<= 0.2.0"] }
+```
+
+# `Sender::send` pointer-as-value transmute causes OOB read and fake-Arc drop under tx/rx race
+
+Maintainer acknowledged on 2026-05-15; v0.3.0 ships the fix plus a regression test (`race_disconnect_does_not_corrupt_sender_or_abort`), and 0.2.0 + all 0.1.x versions have been yanked on crates.io.
+
+`Sender::send` (single-file crate, `src/lib.rs:379-405` in 0.2.0 / commit `23a9ce7`) contains an `unsafe` block in the `DISCONNECTED` arm that transmutes the *pointer to* the producer instead of the producer value:
+
+```rust
+let consumer: spsc::Consumer<T> = std::mem::transmute(self.producer.get());
+```
+
+`self.producer` has type `UnsafeCell<spsc::Producer<T>>`. `UnsafeCell::<X>::get(&self)` returns `*mut X` — a pointer-sized value. The author's intent (per the surrounding comment) was a value-level transmute (`Producer<T>` and `Consumer<T>` are layout-compatible newtypes around `Arc<Buffer<T>>`). The shipped code is one indirection off: it copies 8 bytes of pointer into the bytes of a `Consumer<T>`, whose internal `Arc::ptr` is then the address of the producer field on the `Sender`, not the real `ArcInner<Buffer<T>>`.
+
+The subsequent `consumer.try_pop()` walks `Buffer<T>` fields at offsets that lie inside the `Sender<T>` struct (`send_new`, `inner`) — an OOB read of the Sender's own frame. The fake `Consumer<T>`'s `Drop` then decrements bytes treated as `AtomicUsize::strong_count` and calls `dealloc(...)` on a non-allocated address.
+
+The branch is not reachable single-threaded: the receiver-drop and `connected = false` guard prevent it. The trigger is a TOCTOU race: the sender's `connected.load()` wins but the receiver's `counter.compare_exchange(_, DISCONNECTED, ...)` wins. Under contention the race reproduces in a few hundred trials.
+
+## Trigger
+
+```rust
+use std::thread;
+use unbounded_spsc::channel;
+
+for _ in 0..500 {
+    let (tx, rx) = channel::<Box<u64>>();
+    let h = thread::spawn(move || {
+        for _ in 0..10_000 { let _ = tx.send(Box::new(0xDEAD_BEEF)); }
+    });
+    drop(rx);
+    let _ = h.join();
+}
+// Release: SIGSEGV within a few iterations.
+// ASan: stack-buffer-overflow inside fake-Consumer::try_pop.
+```
+
+## Smoking-gun upstream evidence
+
+`src/lib.rs:975` in the project's own test suite carries:
+
+```
+// TODO: failures
+// - failed with assertion on line 394 in send fn
+//   assert!(second.is_none())
+```
+
+That assertion lives in the same transmute block. The author has observed the symptom — `try_pop()` returning `Some` where logically there should be `None` — as a flaky test rather than recognising it as UB.