1 files changed, 30 insertions, 0 deletions
diff --git a/Documentation/transfer-data-leaks.txt b/Documentation/transfer-data-leaks.txt
new file mode 100644
index 0000000..914bacc
--- /dev/null
+++ b/Documentation/transfer-data-leaks.txt
@@ -0,0 +1,30 @@
+SECURITY
+--------
+The fetch and push protocols are not designed to prevent one side from
+stealing data from the other repository that was not intended to be
+shared. If you have private data that you need to protect from a malicious
+peer, your best option is to store it in another repository. This applies
+to both clients and servers. In particular, namespaces on a server are not
+effective for read access control; you should only grant read access to a
+namespace to clients that you would trust with read access to the entire
+repository.
+
+The known attack vectors are as follows:
+
+. The victim sends "have" lines advertising the IDs of objects it has that
+  are not explicitly intended to be shared but can be used to optimize the
+  transfer if the peer also has them. The attacker chooses an object ID X
+  to steal and sends a ref to X, but isn't required to send the content of
+  X because the victim already has it. Now the victim believes that the
+  attacker has X, and it sends the content of X back to the attacker
+  later. (This attack is most straightforward for a client to perform on a
+  server, by creating a ref to X in the namespace the client has access
+  to and then fetching it. The most likely way for a server to perform it
+  on a client is to "merge" X into a public branch and hope that the user
+  does additional work on this branch and pushes it back to the server
+  without noticing the merge.)
+
+. As in #1, the attacker chooses an object ID X to steal. The victim sends
+  an object Y that the attacker already has, and the attacker falsely
+  claims to have X and not Y, so the victim sends Y as a delta against X.
+  The delta reveals regions of X that are similar to Y to the attacker.