A use-after-free exists in the Zephyr second-generation work queue (kernel/work.c) in the handling of delayable work timeouts. When a delayable work item's timeout has been dequeued and its handler work_timeout() is in flight (blocked acquiring the work-queue spinlock), a concurrent cancellation does not wait for that handler to finish. In unschedule_locked() the pre-fix code called z_abort_timeout(), which for an already-announcing record returns -EINVAL without removing it; cancel_async_locked() then observes the work as idle, so even k_work_cancel_delayable_sync() and k_work_flush_delayable() return without blocking on the in-flight handler. Because those are the APIs the kernel header documents as the safe way to cancel before freeing a k_work_delayable, a caller that frees the object immediately after a successful sync cancel can race the still-pending handler. work_timeout() subsequently dereferences the freed record: it reads to->dticks via z_is_timeout_handler_canceled() and, if the freed slot has been reused so the bail check fails, performs a read-modify-write of wp->flags (K_WORK_DELAYED_BIT) and submits work against a stale dw->queue pointer — a use-after-free read and write. The k_work API is kernel-mode only (no __syscall entry point), so this is a kernel-internal concurrency defect rather than a userspace privilege escalation. Triggering it requires an SMP build and a subsystem that schedules and then frees (or reschedules) a delayable work item in the narrow window while its timeout is announcing; an attacker able to influence the timing of such teardown (for example via connection churn driving subsystem timers) has a plausible but probabilistic path. The impact is kernel memory corruption or crash (denial of service). The fix makes unschedule_locked() wait, by spinning on z_try_abort_timeout() returning -EAGAIN while releasing and re-acquiring the work spinlock, until any in-flight handler completes before returning, and switches work_timeout() to atomic K_WORK_DELAYED_BIT ownership. This closes both the free-then-handler use-after-free and the related reschedule early-fire race.

The user-space system-call verifier z_vrfy_z_log_msg_static_create() in subsys/logging/log_msg.c was a pure pass-through: it forwarded the caller-supplied source, desc, package, and data arguments directly to the kernel-mode implementation z_impl_z_log_msg_static_create() without performing any of the mandatory K_SYSCALL_* checks. Because z_log_msg_static_create() is declared __syscall, under CONFIG_USERSPACE any unprivileged user-mode thread can invoke it directly with fully attacker-controlled arguments. The kernel-mode handler dereferences each of these untrusted values: frontend_runtime_filtering() reads through the source pointer as a struct log_source_dynamic_data, cbprintf_package_copy() reads desc.package_len bytes from the package pointer, and z_log_msg_finalize() performs a memcpy() of desc.data_len bytes from the data pointer. With no verification, a user thread can supply arbitrary kernel addresses and arbitrary lengths, and the kernel will read from them. The impact is a kernel-mode denial of service (the kernel faults dereferencing an attacker-chosen pointer) and, where a log backend output is observable to the attacker, disclosure of arbitrary kernel memory copied into the emitted log message — a confidentiality breach across the user/kernel boundary that the userspace sandbox is meant to enforce. The reads do not corrupt kernel memory, so there is no out-of-bounds write primitive. The fix adds the required validation to the verifier: it bounds desc.package_len against Z_LOG_MSG_MAX_PACKAGE, rejects non-NULL/length mismatches, and applies K_SYSCALL_MEMORY_READ() to package, data, and (when runtime filtering with a frontend is enabled) source, so any out-of-bounds or kernel pointer now raises K_OOPS instead of being honored.

The Bluetooth host GATT client function parse_read_std_char_desc() in subsys/bluetooth/host/gatt.c parses an ATT Read By Type Response received from a remote GATT server during BT_GATT_DISCOVER_STD_CHAR_DESC discovery. The per-entry stride rsp->len is taken directly from the peer's PDU, and the parse loop both tests its exit condition (length >= rsp->len) and advances (length -= rsp->len, pdu += rsp->len) using that value. The minimum value of rsp->len was never validated before the loop. A malicious or malfunctioning peer can reply with rsp->len = 0. Because length is unsigned and never decreases, the loop condition stays true forever and the read pointer never advances; as long as the body is at least a few bytes with a non-zero handle and a matching descriptor UUID, the host repeatedly re-parses the same bytes and invokes the discovery callback, never terminating. This hangs the Bluetooth host processing thread (CWE-835, loop with unreachable exit condition). The condition is reachable by any connected peer once the local device initiates standard-descriptor-value discovery; GATT discovery does not require bonding or encryption, so an unauthenticated adjacent attacker that the device connects to can trigger it. The impact is denial of service of the Bluetooth subsystem (and likely a watchdog reset on constrained targets); there is no memory disclosure or corruption. The fix adds a rsp->len < sizeof(struct bt_att_data) check before the loop, rejecting under-length responses so the stride is always non-zero and the loop terminates. The sibling parsers parse_include() and parse_characteristic() already validated rsp->len and are unaffected.

The USB device-side CDC NCM class control-to-host handler usbd_cdc_ncm_cth in subsys/usb/device_next/class/usbd_cdc_ncm.c builds a fixed-size response for the GET_NTB_PARAMETERS (28-byte struct ntb_parameters) and GET_NTB_INPUT_SIZE (8-byte struct ntb_input_size) class requests and copies the whole structure into the control DATA IN buffer with net_buf_add_mem(buf, ..., sizeof(...)), ignoring the host-supplied wLength. The control DATA IN buffer is allocated by the USB stack with a capacity of exactly wLength bytes (usbd_ep_ctrl_data_in_alloc -> udc_ctrl_data_alloc -> net_buf_alloc_len(&udc_ep_pool, wLength); no round-up is applied for the IN endpoint). Because net_buf_add_mem/net_buf_simple_add only bounds the copy with an __ASSERT_NO_MSG, which is compiled out in production builds, a host that issues one of these standard CDC NCM control requests with a wLength smaller than the response structure (e.g. wLength = 1) causes the handler to memcpy up to 27 bytes past the end of the allocated pool buffer. The request fields come straight from the USB SETUP packet, so any host (or USB interposer) the Zephyr device enumerates against can trigger the overflow with no authentication once an image built with the device_next USB stack and the CDC NCM class is connected. The out-of-bounds write corrupts adjacent allocations and metadata in the shared udc_ep_pool, primarily causing memory corruption and denial of service of the USB stack; the overflow length is bounded (<= 27 bytes) and the written content is fixed device constants, and the bug reads nothing back so there is no information disclosure. The fix clamps the copy with MIN(sizeof(...), setup->wLength), matching the existing CDC ACM handler.

On the Zephyr ARM port, enabling the hardware FPU (CONFIG_FPU) forces the "Floating point ABI" choice, which defaults to CONFIG_FP_HARDABI. Both FP_HARDABI and FP_SOFTABI permit the compiler to emit hardware FP instructions in any function, even code that never uses floating-point types. However, the callee-saved FP registers (s16-s31 / d8-d15) are only saved and restored across a context switch when CONFIG_FPU_SHARING is enabled (arch/arm/core/cortex_m/swap_helper.S and arch/arm/core/cortex_a_r/swap_helper.S), and prior to this fix selecting an ABI did not enable FPU register sharing, which defaults off. In a build that enables the FPU with the default ABI but leaves CONFIG_FPU_SHARING disabled, the kernel preserves no callee-saved FP register state across thread switches. The documented precondition for this "unshared" mode — that only a single thread ever executes FP instructions — is silently violated because the compiler may generate FP instructions in every thread. Under CONFIG_USERSPACE, where threads are mutually isolated, this becomes an information-disclosure boundary crossing: a victim thread can leave secret-derived values in s16-s31, and a co-resident unprivileged thread can read those registers directly (FP register access is not privilege-gated), recovering data left behind by another thread. Without userspace the same defect causes cross-thread FP state corruption (a correctness fault). The leak is bounded to the 16 callee-saved single-precision registers and is opportunistic, so impact is low. The fix makes FP_HARDABI and FP_SOFTABI select CONFIG_FPU_SHARING and tags every thread with K_FP_REGS at creation, so callee-saved FP state is always preserved across context switches whenever the compiler may emit FP instructions.

The Bluetooth HCI driver for Bouffalo Lab on-chip BLE controllers (BL60x/BL70x/BL61x), bt_bflb_send() in drivers/bluetooth/hci/hci_bflb.c, violates the bt_hci_driver_api.send() buffer-ownership contract. That contract (documented at include/zephyr/drivers/bluetooth.h) requires the buffer reference to be consumed only on success; on error the caller still owns the reference and unrefs it. The driver instead routed all error paths through a shared label that unconditionally called net_buf_unref(buf) before returning the error code, consuming the buffer on failure as well. When send() returns an error, the host TX path (send_buf() in subsys/bluetooth/host/conn.c) unrefs the same buffer again, believing it still owns it. This double-unref over-decrements the net_buf reference count. Because the buffer is a TX fragment whose destroy callback also decrements its still-queued parent buffer, the parent is freed prematurely while reachable on the connection TX queue, producing a use-after-free and corruption of the shared net_buf pool rather than a benign leak. The error conditions are on the host-to-controller transmit path (controller send failure, or an unsupported H:4 packet type), so they are not driven directly by attacker-supplied radio bytes; a remote/adjacent peer can influence them only indirectly, e.g. by inducing controller TX failures under heavy link load. The consequence when reached is BLE-stack denial of service (crash / pool corruption) with possible further memory corruption, bounded to devices using one of these Bouffalo Lab on-chip controllers.

The Realtek BEE Bluetooth HCI driver's send callback, bt_hci_bee_send() in drivers/bluetooth/hci/hci_bee.c, violated the bt_hci_driver_api buffer-ownership contract. That contract requires the driver to consume (unref) the transmit net_buf only on success; on an error return the host caller retains ownership and unrefs the buffer itself. The pre-fix code routed all error paths through a shared cleanup label that unconditionally called net_buf_unref(buf) before returning the error code. Because the host TX paths (in subsys/bluetooth/host/hci_core.c) unref the buffer again after send() returns an error, the buffer is freed twice: the driver returns it to its net_buf pool and the host then unrefs the already-freed buffer, corrupting the shared pool / underflowing the reference count (CWE-415). The same error branch additionally dereferenced buf->len inside a LOG_ERR call after the buffer had already been unref'd, a read of freed memory (CWE-416) that is compiled in at the default error log level. The failing edges are reached when the controller's host-to-controller buffer allocation fails or the controller send fails (resource-exhaustion / IO conditions). A remote Bluetooth peer can push the device toward these conditions indirectly by driving heavy host transmit activity, at which point the double-free corrupts the host net_buf pool and most likely crashes the device, with residual potential for further memory corruption. The impact is confined to builds using this specific Realtek BEE HCI driver. The fix returns early from each error path without unreffing and unrefs the buffer only on the success path, restoring the ownership contract and eliminating both the double-free and the use-after-free read.

The USB DFU class implementation in Zephyr's new (experimental) device_next USB device stack contains a NULL pointer dereference in handle_download() (subsys/usb/device_next/class/usbd_dfu.c). The handler computes MIN(setup->wLength, buf->len) and passes buf->data to the image write callback without checking that the buf net_buf pointer is non-NULL. The handler is reached over the USB control endpoint, driven by the USB host. For a DFU_DNLOAD (download) request with no Data OUT stage — notably the zero-length terminating download that the DFU protocol uses to end a firmware transfer — the USB core invokes the class handler with a NULL buffer. After the device has been advanced to the DFU_DNLOAD_IDLE state (by sending one valid download block and a GET_STATUS), a zero-length DFU_DNLOAD reaches handle_download() with buf == NULL, dereferencing it. The result is a NULL+offset read that triggers a fatal CPU fault, i.e. a denial of service (device crash/reset). The attacker is whatever controls the USB host the device is attached to; DFU download support must be enabled with a registered image. There is no memory corruption or information disclosure — impact is limited to availability. The fix adds an explicit if (buf != NULL) guard so the callback receives a zero-length, NULL-data transfer instead of crashing.

The Linkable Loadable Extensions (llext) subsystem mis-handles PLT/RELA relocation entries when linking a relocatable (partially-linked) ELF extension. In llext_link_plt() (subsys/llext/llext_link.c), the relocatable branch (tgt != NULL, the path used for Xtensa relocatable objects) computed the patch address as ext->mem[LLEXT_MEM_TEXT] - text.sh_offset + rela.r_offset + tgt->sh_offset and then performed the relocation write there without validating rela.r_offset. Its sibling shared/dynamic branch already rejected out-of-range offsets via llext_file_offset(). rela.r_offset is read directly from the ELF's RELA table, so a crafted entry with an offset larger than the target section makes the write land arbitrarily far outside the extension's text buffer. The result is an attacker-influenced out-of-bounds write (the location via r_offset, the written value being the resolved symbol address) performed in supervisor context at link time, before any extension code runs. The path is reached from llext_load() whenever an application loads an attacker-influenced ELF extension on Xtensa with writable storage; llext is documented to accept extensions of untrusted origin. Impact is supervisor-context memory corruption (integrity and availability loss, and a sandbox-boundary escape for user-mode extensions). Exploitation is gated by the Xtensa relocatable PLT path and writable storage, and turning the out-of-range write into a useful primitive is non-trivial. The fix adds a bound check rejecting any RELA entry whose r_offset >= tgt->sh_size, mirroring the existing validation in the shared branch.

The userspace syscall verifiers z_vrfy_zsock_sendmsg() and z_vrfy_zsock_recvmsg() in subsys/net/lib/sockets/sockets.c snapshot the caller-supplied struct net_msghdr into a kernel-side copy with k_usermode_from_copy(), but then re-read the still-live user struct for subsequent decisions. The kernel iovec shadow buffer is sized from one read of msg->msg_iovlen, while the population loop is bounded by a second, live read of the same field. Because msg points into ordinary user memory, a cooperating second thread in the same memory domain can inflate msg->msg_iovlen in the window between the sizing read and the loop test (a classic double-fetch / TOCTOU). The population loop then iterates past the number of net_iovec slots actually allocated, writing attacker-influenced iov_base/iov_len values beyond the end of the kernel-heap shadow buffer. The recvmsg verifier has the same defect on both its inbound and result write-back loops. The code is reachable from an unprivileged user thread whenever CONFIG_USERSPACE is enabled and the zsock_sendmsg/zsock_recvmsg syscalls are available. A successful race corrupts kernel-managed heap memory across the user-to-kernel privilege boundary, yielding a local privilege-escalation primitive or, at minimum, a kernel-fault denial of service. The fix copies the header once and derives every size, bound, and gate from the snapshot, copying each iovec entry atomically so its base and length can no longer be raced apart.