diff --git a/crates/bashkit-cpython-wasm/artifacts/MANIFEST b/crates/bashkit-cpython-wasm/artifacts/MANIFEST
index d680129b7..41c3a5fec 100644
--- a/crates/bashkit-cpython-wasm/artifacts/MANIFEST
+++ b/crates/bashkit-cpython-wasm/artifacts/MANIFEST
@@ -3,5 +3,5 @@ wasi_sdk=24
 wasmtime_wizer=49.0.1
 zlib=1.3.1
 sqlite=sqlite-amalgamation-3500400
-ac54edd6cb27e55c00500edc995555a69e3654379c34cda437125b6ccdc694d4  python.wasm.gz
-e4c1d3a9db1f8ad2f8a89900055cc4c515897b8d6d2501cfbb8fc3bc76082b29  python314.zip
+6f4c6d7b93992426af97884dfb603c2ef9ddda9ad20e51e2b79e84c18eb1e140  python.wasm.gz
+672dd03d9e31bee0d92c14f10f772b97c5daedbc34b9f14d5ebc21b086b77daf  python314.zip
diff --git a/crates/bashkit-cpython-wasm/artifacts/python.wasm.gz b/crates/bashkit-cpython-wasm/artifacts/python.wasm.gz
index b80a7ed15..759146588 100644
Binary files a/crates/bashkit-cpython-wasm/artifacts/python.wasm.gz and b/crates/bashkit-cpython-wasm/artifacts/python.wasm.gz differ
diff --git a/crates/bashkit-cpython-wasm/artifacts/python314.zip b/crates/bashkit-cpython-wasm/artifacts/python314.zip
index 5a1a4132d..cdb3084a0 100644
Binary files a/crates/bashkit-cpython-wasm/artifacts/python314.zip and b/crates/bashkit-cpython-wasm/artifacts/python314.zip differ
diff --git a/crates/bashkit-cpython-wasm/guest/_bashkit_boot.py b/crates/bashkit-cpython-wasm/guest/_bashkit_boot.py
index 0833a49d7..3200f13b9 100644
--- a/crates/bashkit-cpython-wasm/guest/_bashkit_boot.py
+++ b/crates/bashkit-cpython-wasm/guest/_bashkit_boot.py
@@ -8,7 +8,10 @@
 Decisions:
 - Tracebacks hide this driver's frames so output matches real `python3`.
 - `random` is re-seeded per call: the snapshot would otherwise hand every
-  tenant the same sequence.
+  tenant the same sequence. Seeding is lazy (first use in a call), since it
+  costs ~0.2 ms on Pulley and most calls never touch `random`. Calling the
+  C base method directly (`_random.Random.random(random._inst)`) skips the
+  seed and sees the snapshot's state; `random` is not a CSPRNG anyway.
 - The snapshot heap is frozen (`gc.freeze`) so garbage collection only
   scans objects created by the current call.
 - Snapshot objects are made immortal before freezing: reference counting
@@ -73,6 +76,7 @@ def preload():
         except Exception:  # pragma: no cover - optional modules
             pass
     _patch_asyncio()
+    _install_lazy_random()
     # Move every object the snapshot holds into the permanent generation.
     # The collector then never traverses (or writes to) snapshot memory, so
     # per-call GC cost scales with what the script allocates, and snapshot
@@ -98,6 +102,63 @@ def _immortalize_snapshot():
     _bashkit.immortalize(gc.get_referents(*objects))
 
 
+_RANDOM_API = (
+    "seed", "random", "uniform", "triangular", "randint", "choice", "randrange",
+    "sample", "shuffle", "choices", "normalvariate", "lognormvariate",
+    "expovariate", "vonmisesvariate", "gammavariate", "gauss", "betavariate",
+    "binomialvariate", "paretovariate", "weibullvariate", "getstate",
+    "setstate", "getrandbits", "randbytes",
+)
+_LAZY_ENTRY_POINTS = ("random", "getrandbits", "getstate", "seed", "setstate")
+
+
+def _install_lazy_random():
+    # The module-level functions are methods bound to `random._inst`. Give
+    # that instance a subclass whose entry points seed it from os.urandom on
+    # first use, then turn it back into a plain Random and rebind the module
+    # functions, so later calls pay nothing extra.
+    import random
+
+    base = random.Random
+
+    def wake(self):
+        # Only the overridden entry points need rebinding; the other module
+        # functions are base-class methods and see the class switch. Module
+        # attribute writes are slow on Pulley (~40 us each), so keep it to five.
+        self.__class__ = base
+        for name in _LAZY_ENTRY_POINTS:
+            setattr(random, name, getattr(self, name))
+
+    class _LazyRandom(base):
+        def random(self):
+            wake(self)
+            base.seed(self)
+            return self.random()
+
+        def getrandbits(self, k):
+            wake(self)
+            base.seed(self)
+            return self.getrandbits(k)
+
+        def getstate(self):
+            wake(self)
+            base.seed(self)
+            return self.getstate()
+
+        def seed(self, *args, **kwargs):
+            wake(self)
+            return self.seed(*args, **kwargs)
+
+        def setstate(self, state):
+            wake(self)
+            return self.setstate(state)
+
+    inst = random._inst
+    inst.__class__ = _LazyRandom
+    for name in _RANDOM_API:
+        setattr(random, name, getattr(inst, name))
+
+
 def _patch_asyncio():
     try:
         from asyncio import selector_events
@@ -197,9 +258,6 @@ def _refresh_process_state():
             os.chdir(pwd)
         except OSError:
             pass
-    random = sys.modules.get("random")
-    if random is not None:
-        random.seed()
 
 
 def _fresh_main(filename=None):
diff --git a/crates/bashkit/docs/cpython.md b/crates/bashkit/docs/cpython.md
index 2e23d1283..4dd7d9665 100644
--- a/crates/bashkit/docs/cpython.md
+++ b/crates/bashkit/docs/cpython.md
@@ -137,7 +137,7 @@ let bash = Bash::builder()
   (`FileNotFoundError`, ...) and messages are correct; code comparing
   `e.errno == errno.ENOENT` works because the `errno` module matches.
 - **Fixed hash seed**: `hash()` of `str`/`bytes` is the same in every call
-  (the seed is baked into the snapshot). `random` is re-seeded per call.
+  (the seed is baked into the snapshot). `random` is re-seeded per call (on first use).
 - **Deep C-level recursion** (for example `repr` of a list nested 100 000
   levels) ends the call with `python3: fatal error: stack overflow in the
   interpreter` instead of `RecursionError`.
diff --git a/crates/bashkit/tests/integration/cpython_integration_tests.rs b/crates/bashkit/tests/integration/cpython_integration_tests.rs
index f8ad3ea62..e77c8b7ac 100644
--- a/crates/bashkit/tests/integration/cpython_integration_tests.rs
+++ b/crates/bashkit/tests/integration/cpython_integration_tests.rs
@@ -566,6 +566,31 @@ async fn random_is_reseeded_per_call() {
     assert_eq!(out, "differ\n");
 }
 
+/// Seeding is lazy: every entry point into the shared `random` instance must
+/// still see a fresh per-call seed, and explicit seeding stays deterministic.
+#[tokio::test]
+async fn lazy_random_seed_covers_every_entry_point() {
+    let differ = |snippet: &str| {
+        format!(
+            "a=$(python3 -c '{snippet}'); b=$(python3 -c '{snippet}'); [ \"$a\" != \"$b\" ] && echo differ"
+        )
+    };
+    for snippet in [
+        "from random import random; print(random())",
+        "import random; print(random.randint(0, 10**18))",
+        "import random; print(random.getrandbits(64))",
+        "import random; print(random.choice(range(10**9)))",
+        "import random; print(random.getstate()[1][:4])",
+        "import random; x = list(range(50)); random.shuffle(x); print(x)",
+    ] {
+        assert_eq!(ok(&differ(snippet)).await, "differ\n", "{snippet}");
+    }
+    let out = ok("python3 -c 'import random; random.seed(42); a = random.random(); random.seed(42); print(a == random.random(), random.Random(42).random() == a)'; \
+         python3 -c 'import random; s = random.getstate(); a = random.random(); random.setstate(s); print(a == random.random())'")
+    .await;
+    assert_eq!(out, "True True\nTrue\n");
+}
+
 #[tokio::test]
 async fn separate_bash_instances_share_nothing() {
     let mut a = bash();
diff --git a/knowledge/log.md b/knowledge/log.md
index 9e104f86c..6d6973d3b 100644
--- a/knowledge/log.md
+++ b/knowledge/log.md
@@ -2,6 +2,7 @@
 
 ## 2026-10-07
 
+* **Performance**: CPython `random` seeds lazily. The snapshot's `random._inst` gets a subclass whose entry points (`random`, `getrandbits`, `getstate`, `seed`, `setstate`) seed from `os.urandom` on first use, then switch it back to `Random` and rebind those five module functions. Calls that never touch `random` skip the reseed (`pass` 3.7 to 3.5 ms). See [CPython WebAssembly Runtime](runtimes/cpython-wasm.md).
 * **Performance**: Recorded CPython benches after #2534/#2535 (`criterion-python-vm-linux-x86_64-1791345434.md`): first call ~19 ms, warm `print(1)` 4.4 ms (was 6.3), 64 parallel sessions 157 ms (was 291), load ~165 calls/s (was ~90). `/benches` now shows Python startup (CPython vs Monty) from these reports. See [Performance Results](operations/performance-results.md).
 * **Feature**: Secret generators. `$SRANDOM` (bash 5.1, 32 bits from `getrandom`), `uuidgen` (v4 only, L-RAND-001) and `openssl rand [-hex|-base64] [-out] N` (only `rand`; capped at 1 MiB, TM-DOS-123). All draw from the OS CSPRNG like `/dev/urandom`, never the `$RANDOM` LCG.
 * **Decision**: CPython stdlib ships as bytecode only. Non-preloaded imports compiled source on Pulley every call (3-7 s each); unchecked-hash `.pyc` cut that ~10x. Unusable modules (`ctypes`, `ssl`, network clients/servers, `pdb`, `pydoc`, `_pyrepl`) and pure-Python C twins dropped so the crate stays under the 10 MiB cap. Dropping preloaded modules was rejected: ~1-2 ms first-call gain, seconds per import lost. See [CPython WebAssembly Runtime](runtimes/cpython-wasm.md), [Limitations](operations/limitations.md) L-CPY-009.
diff --git a/knowledge/runtimes/cpython-wasm.md b/knowledge/runtimes/cpython-wasm.md
index e7556bfb6..a1298884a 100644
--- a/knowledge/runtimes/cpython-wasm.md
+++ b/knowledge/runtimes/cpython-wasm.md
@@ -138,7 +138,9 @@ option parsing (`-c`, `-m`, file, directory, `-`, `-x`, `-W`, `-V`/`-VV`,
 `-h`, ignored init-only flags), fresh `__main__` per call, `sys.argv` and
 `sys.path[0]` rules, tracebacks with driver frames stripped, `SystemExit`
 semantics, `atexit`, then clearing `__main__` + `gc.collect()` so unclosed
-files flush. `random` is re-seeded per call. asyncio's self-pipe is disabled
+files flush. `random` is re-seeded per call, lazily: the module instance
+seeds from `os.urandom` on first use in a call (~0.2 ms saved on calls that
+never use it). asyncio's self-pipe is disabled
 (no sockets, threads or signals exist to need it).
 
 ## Testing