| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| The user-mode syscall verifiers z_vrfy_fuel_gauge_get_props() and z_vrfy_fuel_gauge_set_props() in drivers/fuel_gauge/fuel_gauge_syscall_handlers.c declared two variable-length arrays, union fuel_gauge_prop_val k_vals[len] and fuel_gauge_prop_t k_props[len], sized directly by the caller-supplied len argument. len is an unvalidated size_t taken straight from the syscall ABI, and the VLAs were allocated before any check at all — including before the K_SYSCALL_DRIVER_FUEL_GAUGE() object-permission check. The subsequent k_usermode_from_copy() calls validated only that the user source buffer was readable; the kernel destination was never bounds-checked, since it was sized by the same attacker-chosen len.
Any thread running in user mode with CONFIG_USERSPACE enabled can invoke fuel_gauge_get_props() or fuel_gauge_set_props() with a large len. This first displaces the supervisor stack pointer by an arbitrary attacker-chosen amount — Zephyr does not build with stack-clash probing, so the displacement itself does not fault, and the nested calls made by the verifier then write frames below the privileged stack. If the caller has been granted access to a fuel-gauge device object, the memcpy inside k_usermode_from_copy() additionally writes len * sizeof(union fuel_gauge_prop_val) bytes of fully attacker-controlled data starting well below the stack base. CONFIG_PRIVILEGED_STACK_SIZE defaults to 1024 bytes, so a len of roughly 170 already exhausts it.
The result is an out-of-bounds write in supervisor mode with attacker-controlled length and, on the permitted path, attacker-controlled content — a break out of the user-mode sandbox into kernel memory, leading to kernel code execution or a system crash. A stack guard region does not contain it, because the copy begins below the guard and walks upward, corrupting unprotected memory before the guard is reached. The fix removes the kernel-side copies entirely and validates the caller's arrays in place with K_SYSCALL_MEMORY_ARRAY_READ() / K_SYSCALL_MEMORY_ARRAY_WRITE(), which also handle the len * size multiplication overflow; this is safe because neither fuel_gauge_prop_t nor union fuel_gauge_prop_val contains embedded pointers. |
| net_route_ipv6_packet() in subsys/net/ip/route_ipv6.c resolved the nexthop's link-layer address with net_nbr_get_lladdr(nbr->idx) without first checking whether the neighbor cache entry actually had a linked link-layer address. An unresolved neighbor carries idx == NET_NBR_LLADDR_UNKNOWN (0xff), and net_nbr_get_lladdr() in subsys/net/ip/nbr.c performs no runtime bounds check beyond a NET_ASSERT, returning &net_neighbor_lladdr[255] — roughly 2.5 KB past the end of an array whose default size is CONFIG_NET_IPV6_MAX_NEIGHBORS (8). Because the returned pointer is never NULL, the following lladdr == NULL guard does not catch it.
The function is reached from ipv6_route_packet() in subsys/net/ip/ipv6.c for every received unicast IPv6 packet whose destination is not a local address on the receiving interface; CONFIG_NET_IPV6_ROUTE is enabled by default whenever the IPv6 neighbor cache is, so no router or forwarding configuration is needed. net_route_ipv6_get_info() returns the packet's destination itself as the nexthop when a neighbor cache entry for it exists, and the cache lookup does not skip INCOMPLETE entries. An unauthenticated attacker on the same link can therefore force the unresolved state — for example by eliciting traffic to a spoofed, non-existent neighbor address so that net_ipv6_send_ns() creates an INCOMPLETE entry, or by sending a Router Advertisement with no source link-layer address option, which creates a persistently unresolved router neighbor — and then send a packet addressed to that neighbor.
The result is an out-of-bounds read at a fixed index past the neighbor link-layer address array. On builds with CONFIG_ASSERT enabled the assertion fires and the device panics, giving a repeatable remote denial of service. With assertions disabled, the stale out-of-bounds struct net_linkaddr drives a memcmp() over an attacker-uninfluenced length and, when its len byte passes the NET_LINK_ADDR_MAX_LENGTH check, up to 8 bytes of unrelated static RAM are copied into the outgoing frame's destination link-layer address and transmitted on the link, disclosing them to any listener. There is no out-of-bounds write and the offset is not attacker-controlled, which bounds the impact. |
| The RFC 6528 initial-sequence-number implementation in subsys/net/ip/tcp.c derived every TCP ISN from SHA-256(unique_key || four-tuple) plus a uptime-derived offset, where unique_key is a 128-bit secret filled once by sys_csrand_get(). The return value of that call was discarded and the once guard was latched to true even when the call failed. Because sys_csrand_get() leaves the destination buffer untouched on failure (it deliberately propagates the entropy-driver error rather than filling the buffer), a single failed call left unique_key as its all-zero BSS content for the remainder of the boot, with no retry, no log message and no fallback.
A failure of the cryptographic random source is required to reach the weak state — for example -ENODEV/-EIO from the entropy driver in subsys/random/random_entropy_device.c (the in-tree comment notes that the hardware RNG "might still be gathering entropy during early boot situations"), or psa_generate_random() failing in subsys/random/random_psa.c when PSA crypto is not initialised or is backed by a Bluetooth-HCI entropy device that is not yet up. The first TCP connection is what triggers key generation, so a remote peer that reaches a listening port immediately after boot, or that can provoke a reboot, has indirect influence over whether generation coincides with that window. tcp_init_isn() is called for every passive open in tcp_conn_new() and every active open in net_tcp_connect(), so the poisoned key governs all TCP connections for that boot.
With an all-zero key the ISN becomes a public function of the connection four-tuple plus a device-wide time offset. An off-path attacker can compute the hash term offline for any four-tuple and recover the shared time offset from a single observed ISN, after which the ISN the device will pick for other four-tuples is predictable. That defeats exactly the protection RFC 6528 provides: blind TCP connection spoofing against peers that trust the source address, and blind data injection into or reset of connections whose four-tuple can be guessed. Devices whose entropy source never errors were never in the weak state.
The fix moves key generation into a single guarded helper that latches only on success, logs the error otherwise, and makes tcp_init_isn() fall back to sys_rand32_get() — the behaviour already used when CONFIG_NET_TCP_ISN_RFC6528 is disabled — instead of hashing with a known-constant key. |
| The zbus proxy agent IPC backend in subsys/zbus/proxy_agent/zbus_proxy_agent_ipc.c logged the channel name of a rejected inter-domain frame with a plain %s conversion. The frame type struct zbus_proxy_msg carries a fixed-size channel_name[] array as its last member, and nothing in the transport guarantees the array is NUL-terminated. The only code that verifies termination is zbus_proxy_agent_receive_cb() in subsys/zbus/proxy_agent/zbus_proxy_agent.c, which rejects the frame in precisely those cases — so the warning printed a non-terminated buffer exactly on the error paths where the name had been found invalid (or, for an invalid message_size, had not been inspected at all).
Any peer domain able to place a frame of sizeof(struct zbus_proxy_msg) bytes on the bound ipc_service endpoint can trigger it, by sending a frame with an out-of-range message_size or with channel_name[] containing no NUL byte. Reaching the code requires CONFIG_ZBUS_PROXY_AGENT_IPC and logging built at warning level or above (the default), and requires control over the firmware of the peer domain — typically a second core on the same SoC.
The resulting strlen() inside the log packager walks past the end of the frame object until it finds a zero byte. With the icmsg backend the frame lives in a stack buffer of the IPC work-queue thread, so bytes of that thread's stack are rendered into the log message; with the rpmsg backends the scan continues through the shared vring memory. Impact is bounded to disclosure of a small amount of adjacent memory into the receiving domain's log sink, plus a possible fatal fault if the scan leaves a mapped region; the log packager's own -ENOSPC bound prevents the overrun from becoming a write. The fix bounds the conversion with %.*s and MIN(msg->channel_name_len, sizeof(msg->channel_name)). |
| i2s_esp32_trigger_check() in drivers/i2s/i2s_esp32.c validates the requested direction only for I2S_DIR_BOTH. The I2S_DIR_RX and I2S_DIR_TX branches read dev_cfg->rx.data->configured / dev_cfg->tx.data->configured without first checking the stream pointers. The device instantiation macro I2S_ESP32_STREAM_INIT() sets both .conf and .data to NULL for a direction the devicetree does not describe, so on an instance that wires only one direction — the normal shape for audio output or a worldsemi,ws2812-i2s LED strip — the other direction dereferences a NULL pointer instead of returning an error.
i2s_trigger() is a Zephyr syscall, and z_vrfy_i2s_trigger() in drivers/i2s/i2s_handlers.c validates only the device object and the presence of the trigger API pointer; the dir argument is passed to the driver unvalidated. On a build with CONFIG_USERSPACE enabled, a user-mode thread that has been granted the I2S device can issue a single i2s_trigger() call naming the unwired direction and cause a load from address 0 in kernel mode. Among the Espressif parts that carry this driver, userspace is available in-tree only on RISC-V SoCs with CONFIG_RISCV_PMP, and v4.4.0 is the first release where that is buildable: the ESP32-C6 HPCORE selects RISCV_PMP when it is not built for MCUboot. ESP32-C5 in v4.4.x carries the same PMP-region and userspace linker support but does not select RISCV_PMP by default. Espressif Xtensa targets do not support Zephyr userspace, and in a non-userspace build the bad direction can only come from in-kernel application code.
The impact is limited to availability: the access is a read at offset 0 of the missing stream structure, so there is no attacker-controlled offset, no write primitive and no information disclosure. With the default fatal-error handler the resulting exception halts the system, giving an unprivileged user-mode thread a system-wide denial of service. The fix adds the same pointer check the I2S_DIR_BOTH branch already performed and returns -ENOSYS for a direction the instance does not implement; the driver's other entry points (i2s_esp32_config_check(), i2s_esp32_config_get(), i2s_esp32_read(), i2s_esp32_write()) already guarded the pointers, and a static audit found no equivalent unguarded path.
The driver defect is older than the affected range. The unguarded dereference is present from v4.2.0 (reached through i2s_esp32_trigger_stream(), whose if (stream) guard tests the address of a struct member and is never false) and takes its present i2s_esp32_trigger_check() form in v4.3.0. No in-tree Espressif configuration before v4.4.0 can run a user-mode thread, so in v4.2.x and v4.3.x the direction argument can only come from trusted kernel code. Those releases carry the bug but are not listed as affected; the fix has also been merged to v4.3-branch as hardening. |
| The Bluetooth LE host queues received L2CAP connection-oriented channel (CoC) data for deferred processing through a struct k_work embedded in the channel object (le_chan->rx_work, handler l2cap_rx_process()) whenever the channel uses a dynamic PSM (0x0080-0x00FF). Channel teardown in l2cap_chan_destroy() in subsys/bluetooth/host/l2cap.c cancels the retransmission-timeout work and drains the RX FIFO, but never cancels rx_work. Because that work item was submitted to the system workqueue while HCI receive processing runs on the dedicated Bluetooth RX workqueue (CONFIG_BT_RECV_WORKQ_BT, the default), a queued rx_work item can outlive the channel it points into.
A remote, unauthenticated peer with an established CoC channel triggers this by sending a data K-frame immediately followed by an L2CAP Disconnect Request. The K-frame submits rx_work to the system workqueue; because both workqueue threads are cooperative and the Bluetooth RX workqueue runs at the higher priority (K_PRIO_COOP(CONFIG_BT_RX_PRIO) versus CONFIG_SYSTEM_WORKQUEUE_PRIORITY), the pending item cannot run before the following Disconnect Request is processed in le_disconn_req() -> l2cap_chan_del() -> l2cap_chan_destroy(). The stack then invokes the released() callback, which the API documents as meaning the stack has dropped all references and the application may free the channel memory.
The application therefore frees or re-accepts into an object that the system workqueue still holds in its pending list. If the memory is freed and reallocated, the workqueue later dereferences a list node and a handler function pointer read from reused memory; if the object is re-used for a later connection, l2cap_chan_add() calls k_work_init() on a still-enqueued work item, corrupting the workqueue's pending list so that unrelated work items are dropped or the queue spins on a looped list. A related variant lets l2cap_rx_process() run concurrently with teardown, racing the net_buf unref of le_chan->_sdu and the clearing of chan->conn.
The fix routes the channel RX work to the Bluetooth workqueue - the same context in which every teardown path runs - and adds an explicit k_work_cancel() of le_chan->rx_work in l2cap_chan_destroy(), so no reference to the channel survives the released() callback. Configurations without CONFIG_BT_L2CAP_DYNAMIC_CHANNEL, or that only use SIG-assigned PSMs (EATT 0x0027, OTS 0x0025) which take the inline receive path, are not affected. |
| The Bluetooth Link Layer Control Procedure (LLCP) implementation for Connected Isochronous Stream (CIS) creation retains an RX node (ctx->node_ref.rx, marked NODE_RX_TYPE_RETAIN) so it can later be reused as the host notification — on the peripheral while awaiting the Host's reply to an LL_CIS_REQ, and on the central for the whole duration of a locally initiated CIS Create. In subsys/bluetooth/controller/ll_sw/ull_llcp_cc.c, the "invalid PDU received" paths of llcp_rp_cc_rx() and llcp_lp_cc_rx() terminated the connection and completed the procedure without releasing that retained node, breaking the invariant checked in llcp_lr_check_done() and llcp_rr_check_done() and orphaning the node's memory.
A peer device within radio range can reach this with a single extra LL Control PDU on an unauthenticated, unencrypted ACL link. Against a peripheral, the attacker sends a valid LL_CIS_REQ and then, before the Host replies, any unrelated LL Control PDU (for example LL_VERSION_IND), which ull_cp_rx() routes into the active remote procedure. Against a central performing a CIS Create, a malicious peripheral answers with LL_UNKNOWN_RSP for CIS_REQ, which is dispatched into the active local procedure. No pairing, encryption or user interaction is required; the code is compiled in when CONFIG_BT_CTLR_PERIPHERAL_ISO or CONFIG_BT_CTLR_CENTRAL_ISO is enabled.
In default builds (CONFIG_BT_CTLR_ASSERT_DEBUG is default y) the retained-node assertion fires immediately, producing a controller fatal error and, typically, a system reset from one injected PDU. With the development assertions disabled, each attempt permanently loses one node from the controller's small LL notification pool (LL_PDU_RX_CNT, 2 * CONFIG_BT_CTLR_LLCP_CONN) together with its memq_link_t; repeating the connect-attack-reconnect cycle exhausts the pool, after which notification allocation always fails, RX flow control stalls, and the non-disableable LL_ASSERT_ERR() in llcp_lp_cc_flush() faults. The impact is limited to availability — the leaked node is orphaned, never reused or double-freed — and recovery requires a reboot. |
| The Bluetooth Link Layer control procedure code in subsys/bluetooth/controller/ll_sw/ull_llcp_conn_upd.c retains the received RX node while a Connection Update / Connection Parameter procedure waits for its instant, so the node can later carry the host notification (llcp_rx_node_retain(), which marks it NODE_RX_TYPE_RETAIN and thereby suppresses the normal recycling in ull.c). The default: arm of llcp_rp_cu_rx() and llcp_lp_cu_rx() — the "invalid PDU, terminate the connection" path — completed the procedure without releasing that retained node. llcp_rr_check_done() then dequeued and freed the procedure context with ctx->node_ref.rx still pointing at the retained node, dropping the last reference to it.
A peer device in radio range can reach this without pairing, bonding, or encryption. Against a peripheral, the peer sends a well-formed LL_CONNECTION_UPDATE_IND with an instant a few connection events in the future (the node is retained and the procedure enters RP_CU_STATE_WAIT_INSTANT), then, before the instant is reached, sends any other LL Control PDU such as LL_LENGTH_REQ. ull_cp_rx() routes that PDU into the active remote Connection Update procedure, which takes the invalid-PDU path. The central role is reachable symmetrically after accepting an LL_CONNECTION_PARAM_REQ, and the local-procedure variant is reachable with LL_REJECT_IND.
With CONFIG_BT_CTLR_ASSERT_DEBUG enabled (its default), the resulting state violates the invariant asserted in llcp_rr_check_done(), so the two-PDU sequence produces an immediate fatal error in the controller. With those asserts disabled, each occurrence permanently loses one node from the controller's small fixed RX pool (sized from CONFIG_BT_CTLR_RX_BUFFERS, which defaults to 1); repeating the sequence across reconnections exhausts the pool, after which the link-layer receive path operates on a NULL node. The impact is an unauthenticated, remotely triggerable denial of service persisting until reboot; there is no memory-disclosure or memory-corruption consequence, since the leaked node simply becomes unreachable. |
| The Link Layer Control Procedure (LLCP) implementation of the Zephyr software Bluetooth LE Controller retains the receive node that carried an accepted LL_PHY_UPDATE_IND so that it can later be reused for the host notification when the update instant is reached (llcp_rx_node_retain() in subsys/bluetooth/controller/ll_sw/ull_llcp.c, and the node is deliberately not recycled while marked NODE_RX_TYPE_RETAIN). The invalid-PDU arms of llcp_lp_pu_rx() and llcp_rp_pu_rx() in subsys/bluetooth/controller/ll_sw/ull_llcp_phy.c completed the procedure via llcp_lr_complete() / llcp_rr_complete() without first releasing that retained node, so the procedure context — the only remaining reference to the node — was freed while the node was still held out of the receive pool.
A peer device on an established LE connection can drive this deterministically and without pairing or encryption. Against a peripheral it sends LL_PHY_REQ, receives LL_PHY_RSP, sends a valid LL_PHY_UPDATE_IND with an instant a few connection events in the future (so the node becomes retained), and then, before the instant is reached, sends any other LL Control PDU such as LL_LENGTH_REQ; ull_cp_rx() routes it to the active remote PHY Update procedure, which takes the invalid-PDU path. The mirror case applies to a locally initiated PHY Update followed by an LL_REJECT_IND.
In the default configuration (CONFIG_BT_ASSERT and CONFIG_BT_CTLR_ASSERT_DEBUG both default y) the violated invariant in llcp_lr_check_done() / llcp_rr_check_done() triggers a controller assertion, ending in k_oops() (or k_panic()) — a single crafted PDU sequence from radio range faults the device. With those assertions compiled out, each attempt silently leaks one receive PDU node and its memq link; because the controller receive pool is small (PDU_RX_CNT, driven by CONFIG_BT_CTLR_RX_BUFFERS, which defaults to 1) and each attempt costs the attacker only a reconnect, a few repetitions exhaust the pool and leave Bluetooth inoperable until reboot. On releases v3.4.0 through v3.7.x the assertion is never reached, whatever the configuration, so every attempt leaks silently.
The impact is limited to availability: the orphaned node leaves no dangling pointer that is later dereferenced and is never delivered to the host, so there is no memory corruption or information disclosure. The same pull request applies the identical release to the Connection Update and CIS-create procedures, whose invalid-PDU arms had the same omission. |
| The Goodix GT9xx input driver in drivers/input/input_gt911.c reads the touch point count from the controller's status register and masks it with GT911_TOUCH_POINTS_MSK (0x0F), yielding a value of 0..15. In gt911_process() that value is used directly as the loop bound for filling point_reg[], a stack array sized to CONFIG_INPUT_GT911_MAX_TOUCH_POINTS, whose Kconfig range is 1..5 with a default of 1. No other check constrains the count; the driver relied only on a comment asserting that the controller had been programmed at init to report no more points than configured.
Each loop iteration issues an I2C read of eight bytes straight into point_reg[i], so a controller that reports more points than the array holds causes up to 112 bytes of peer-supplied data to be written past the end of the array, over the stack frame of gt911_process() in the system workqueue thread. Two further loops then read out of bounds from the same array. Triggering it requires control of, or the ability to substitute, the I2C touch controller — plausible on the many supported boards where the GT9xx panel is a pluggable display module or shield rather than an on-PCB part; a spoofed device need only answer the init probes with a supported product ID and a checksum-valid config blob. The same overflow can also occur non-adversarially, with a GT9271-class panel that ignores the driver's touch-count programming and reports up to ten points, or with bus corruption of the single status byte.
The impact is an out-of-bounds stack write with fully attacker-chosen content executing at kernel privilege, i.e. potential control-flow hijack on the host MCU, in addition to out-of-bounds reads and crashes. The attack vector is physical/local hardware access only; there is no network, USB, or syscall path to the defect. The fix clamps the reported count with min() against CONFIG_INPUT_GT911_MAX_TOUCH_POINTS before any array indexing. |
| The NXP MCUX TRNG entropy driver in drivers/entropy/entropy_mcux_trng.c passed the caller's byte count straight to the vendor SDK routine TRNG_GetRandomData(). On i.MX RT5xx and RT6xx parts the SDK compiles its TRNG_SW_HEALTH_TESTS variant, which always copies whole 32-bit words and draws entropy rounded up to a multiple of 128 bytes. Its "caller buffer is full" guard tests dataSize == 0, so a request whose length is not a multiple of four makes dataSize underflow past zero and the SDK keeps writing into the caller's buffer for the entire extraction: a 1-byte request results in 128 bytes written, and any non-word-multiple length overflows by up to 127 bytes.
entropy_get_entropy() is a syscall, and its verifier in drivers/entropy/entropy_handlers.c validates only the requested length via K_SYSCALL_MEMORY_WRITE(). With CONFIG_USERSPACE enabled, an unprivileged user-mode thread that has merely been granted the entropy device can therefore choose both the destination address and a length such as 1, and cause the kernel to write up to 127 bytes beyond the region it proved it owns. On these Cortex-M33 targets there is no MMU, so user partitions and kernel data share one SRAM and the overflow can land in adjacent kernel state. The same defect is reached from kernel mode by any caller requesting a non-word-multiple length, including getentropy() and, in builds where sys_csrand_get()/sys_rand_get() resolve to the hardware generator, sys_rand8_get() and sys_rand16_get().
Impact is memory corruption of up to 127 bytes immediately following the supplied buffer — typically the caller's stack in kernel-mode use, or memory outside the caller's partition when driven through the syscall. The overflow offset is fully determined by the requested length and is therefore deterministic, while the written content is uncontrolled TRNG output; the practical consequences range from crashes and unpredictable state corruption to opportunistic escalation when kernel bookkeeping such as object permission bitmaps is overwritten. The affected devices are those where the MCUX SDK enables TRNG_SW_HEALTH_TESTS (MIMXRT595S, MIMXRT555S, MIMXRT533S, MIMXRT685S, MIMXRT633S), on which the TRNG is the zephyr,entropy chosen node; other SoCs using this driver take the SDK path that clamps the copy size and are unaffected.
The fix routes any unaligned prefix and any sub-word tail through a local bounce word and hands the SDK only word-multiple sizes, so the SDK's word-granular writes can no longer pass the end of the caller's buffer. |
| The user-mode verification handler for the device_deinit() system call, z_vrfy_device_deinit() in kernel/device.c, validated its dev argument with K_SYSCALL_OBJ_INIT(dev, K_OBJ_ANY). k_object_validate() short-circuits its type comparison when the requested type is K_OBJ_ANY, so the check reduced to "this pointer is the base address of some kernel object the calling thread has been granted" — the object's actual type was never compared, and K_SYSCALL_OBJ_INIT also skips the initialization-state check. The sibling handlers z_vrfy_device_init() and z_vrfy_device_is_ready() already used K_OBJ_DRIVER_ANY and were unaffected.
A thread running in user mode can therefore pass any kernel object it holds permission on — most usefully a thread stack object obtained from the k_thread_stack_alloc() syscall or a statically defined K_THREAD_STACK it was granted in order to spawn a child user thread — whose backing memory is writable from user mode. z_impl_device_deinit() then interprets those attacker-written bytes as a struct device: it dereferences the state pointer read out of the object, calls the function pointer read out of ops.deinit, and on success writes through state again. The result is an indirect call to an arbitrary address executed in supervisor mode, plus an arbitrary kernel read and a single-byte kernel write.
Exploitation gives a local unprivileged thread full kernel code execution, defeating the CONFIG_USERSPACE isolation boundary entirely; a less precise attempt yields a supervisor-mode fault and a system crash. The defect is only reachable in builds that enable both CONFIG_USERSPACE and CONFIG_DEVICE_DEINIT_SUPPORT — with de-initialization support disabled, z_impl_device_deinit() returns -ENOTSUP without ever dereferencing the pointer. In v4.2.x and v4.3.x, CONFIG_DEVICE_DEINIT_SUPPORT defaulted to y, so every CONFIG_USERSPACE build of those releases is exposed unless the option was explicitly turned off. From v4.4.0 the option is opt-in (no default, and not selected by any in-tree subsystem), so a v4.4.x build is exposed only if it enables the option explicitly. The v4.2 line is no longer maintained and receives no backport.
The fix changes the object check to K_OBJ_DRIVER_ANY, which constrains the argument to the build-generated driver object type range (K_OBJ_DRIVER_FIRST..K_OBJ_DRIVER_LAST) — the real struct device instances placed by the linker — so the state and ops.deinit fields are once again kernel-controlled. |
| The ARM64 MMU back-end allocated address space identifiers (ASIDs) for memory domains with a bare round-robin counter in arch_mem_domain_init() (arch/arm64/core/mmu.c). VM_ASID_BITS is 8, so only 255 ASIDs exist; once the counter wrapped, arch_mem_domain_init() could hand an ASID to a new domain while a still-live domain held the same one. Domain-private mappings are installed non-global (MT_NG), so the ASID is the only tag separating one domain's cached translations from another's in the TLB.
The context-switch path in z_arm64_swap_ptables() only flushes the TLB when the outgoing and incoming domains carry the same ASID, which does not cover a duplicate reached through a third domain: for domains A and C sharing an ASID and an unrelated domain B, the schedule A -> B -> C never takes the flush branch, so the ASID-tagged entries A populated remain resident while C runs. Under SMP two live domains sharing an ASID can additionally be resident on two CPUs at once, which the architecture does not allow for distinct translation-table sets.
Triggering the wrap requires a CONFIG_USERSPACE application on ARM64 that creates more than 255 memory domains over its lifetime; k_mem_domain_init() and k_mem_domain_deinit() are supervisor-only APIs and are not exposed as syscalls, so an unprivileged thread cannot drive the counter directly. Once two live domains alias, however, a user-mode thread in one domain can read and write memory belonging to the other domain's partitions and thread stacks with that domain's permissions, defeating the memory-domain isolation boundary.
The fix scans the live domain_list before assigning an ASID, advances the round-robin counter past ASIDs already in use, and returns -ENOMEM when all are taken, so domain creation fails closed instead of silently aliasing. |
| The ITE IT8xxx2 SHI host-command backend (subsys/mgmt/ec_host_cmd/backends/ec_host_cmd_backend_shi_ite.c) copied the 8-byte host-command request header from the SPI Rx FIFO directly into the shared receive buffer data->in_msg and only afterwards checked the protocol version and the derived packet length. The interrupt handler also accepted a chip-select assertion and an Rx-valid-length (RVLI) interrupt in any driver state other than SHI_STATE_DISABLED, so a new header could be parsed while the host-command thread was still processing the previous request out of the very same buffer.
The host processor is the SPI controller and drives both chip select and the clock. After sending a well-formed request it can immediately de-assert chip select — which returns the driver to the ready state and re-enables the FIFO — and start a second transaction carrying a header with data_len = 0xFFFF. Those eight bytes are written into in_msg before the oversized length is rejected, so they land in a buffer whose contents verify_rx() in subsys/mgmt/ec_host_cmd/ec_host_cmd_handler.c has already validated. If this lands in the window before the host-command thread executes args.input_buf_size = rx_header->data_len, the framework hands the registered command handler a 65535-byte input length over a 256-byte buffer.
The result is an out-of-bounds read of up to roughly 64 KiB beyond the request buffer: command handlers that copy or echo input_buf_size bytes disclose adjacent embedded-controller memory back to the host or overflow the response buffer, and a read past the end of SRAM faults the controller. The same race also allows cmd_id and cmd_ver to be swapped after checksum verification and after handler lookup. Exploitation requires the ability to drive the inter-processor SHI bus (a compromised host OS or physical access to the SPI lines) and winning a timing race, which the SPI controller can retry indefinitely.
The fix parses the header into a local struct ec_host_cmd_request_header and copies it into in_msg only after the length has been bounded by sizeof(data->in_msg), and ignores chip-select and RVLI interrupts outside SHI_STATE_READY_TO_RECV/SHI_STATE_RECEIVING. A residual, bounded race remains: an end-of-transaction interrupt still resets the state to ready while the host-command thread owns the buffer, so a valid second request can still overwrite the in-flight request's contents, unlike the NPCX backend which parks in SHI_STATE_CNL_RESP_NOT_RDY while the buffer is in use. |
| The LE Audio Broadcast Sink in subsys/bluetooth/audio/bap_broadcast_sink.c copies subgroup metadata from a received Basic Audio Announcement (BASE) into the static Broadcast Audio Scan Service parameter structure mod_src_param without any bounds check. In base_subgroup_meta_cb() the destination element was selected as mod_src_param.subgroups[mod_src_param.num_subgroups] with no test against ARRAY_SIZE(mod_src_param.subgroups) (sized by CONFIG_BT_BAP_BASS_MAX_SUBGROUPS, default 1), and the metadata was copied with memcpy() using the raw on-air length returned by bt_bap_base_get_subgroup_codec_meta() into a metadata array sized by CONFIG_BT_AUDIO_CODEC_CFG_MAX_METADATA_SIZE (default 4). The BASE validator bt_bap_base_get_base_from_ad() only checks structural consistency and permits up to ~24 subgroups and metadata LTVs of ~240 octets.
The defect is reached from the periodic advertising receive callback: pa_recv() → bt_data_parse() → pa_decode_base() → update_recv_state_base() → bt_bap_base_foreach_subgroup() → base_subgroup_meta_cb(). Every broadcast sink registers a scan-delegator receive state at creation (bt_bap_broadcast_sink_create() calls broadcast_sink_add_src()), and CONFIG_BT_BAP_BROADCAST_SINK depends on CONFIG_BT_BAP_SCAN_DELEGATOR, so the path is active in every broadcast-sink build once the device is periodic-advertising-synced. An attacker in radio range who operates a broadcast source the device syncs to — or who impersonates the advertiser address and SID of one already in use, periodic advertising data being unauthenticated — can change the BASE at will; each new BASE is re-parsed.
A crafted BASE therefore writes attacker-chosen bytes past the end of a fixed static object in .bss: up to roughly 236 bytes for an oversized metadata LTV, plus whole struct bt_bap_bass_subgroup records for each subgroup beyond CONFIG_BT_BAP_BASS_MAX_SUBGROUPS. This is memory corruption of adjacent Bluetooth-audio state reachable with no pairing, bonding or GATT connection, with a potential for remote code execution in the Bluetooth RX thread; in addition, the unvalidated metadata_len is forwarded to bt_bap_scan_delegator_mod_src(), which neither clamps it nor rejects it, leading to a further copy into the receive state and to out-of-bounds memory being disclosed in the BASS receive-state notification sent to a connected Broadcast Assistant.
The fix rejects a BASE carrying more subgroups than the receive state can hold (discarding the update entirely) and omits metadata that does not fit rather than copying it, and additionally honours the previously-ignored error return of the subgroup decode pass. |
| dynamic_object_create() in kernel/userspace/userspace.c computed the backing allocation for a dynamically allocated kernel object as obj_size_get(otype) + size, and for thread stack elements as STACK_ELEMENT_DATA_SIZE(size) (a round-up plus fixed overhead), without checking either expression for unsigned wrap-around. A size close to SIZE_MAX makes the computed total wrap to a very small value, so the heap chunk handed out is a few bytes while the object descriptor is still tagged with the full requested type and registered in the kernel object table.
The size argument reaches that arithmetic directly from user mode. k_object_alloc_size() is declared __syscall in include/zephyr/sys/kobject.h, its verifier z_vrfy_k_object_alloc_size() in kernel/userspace/userspace_handler.c is a bare pass-through, and z_object_alloc() only range-checks otype — nothing bounds size. The stack-element branch is additionally reachable through the k_thread_stack_alloc() syscall via kernel/dynamic.c. Because subsequent kernel-object validation checks only the object's type and initialization state, the undersized handle passes K_SYSCALL_OBJ_INIT()/K_SYSCALL_OBJ_NEVER_INIT(), and the matching init syscall (for example k_mutex_init(), k_sem_init(), or k_thread_create()) then writes a complete object over the truncated allocation.
An unprivileged user-mode thread can therefore trigger a supervisor-mode out-of-bounds write into the kernel resource-pool heap, of a size and content it substantially controls, corrupting sys_heap chunk metadata and adjacent kernel objects. Under CONFIG_GEN_PRIV_STACKS the thread-stack branch additionally stores an attacker-influenced wild pointer as a user thread's privileged stack base. The practical result is escape from the CONFIG_USERSPACE sandbox — kernel-level code execution or at minimum kernel memory corruption and system compromise.
Exploitation requires CONFIG_USERSPACE together with CONFIG_DYNAMIC_OBJECTS (also selected by CONFIG_DYNAMIC_THREAD under userspace), and a calling thread with an assigned resource pool. The fix rejects both overflowing computations and frees the partially built descriptor. |
| ieee802154_decipher_data_frame() in subsys/net/l2/ieee802154/ieee802154_frame.c computed payload_len = net_pkt_get_len(pkt) - ll_hdr_len - authtag_len without first checking that the received frame is at least ll_hdr_len + authtag_len bytes long. All three variables are uint8_t, so a frame whose payload is shorter than the configured authentication tag makes the subtraction wrap around to a large value (up to 255).
The wrapped length is passed unchanged to ieee802154_decrypt_auth() and on to the CCM operation as cipher_pkt.in_len/out_buf_max, with apkt->tag pointing at frame + ll_hdr_len + payload_len. Because the receive buffer is allocated to the exact length of the frame received from the radio driver, the crypto layer then reads several hundred bytes past the end of the packet buffer and writes the same number of decrypted bytes back over it in place. The frame's authentication tag is only verified after this processing has taken place, so no key material, association or prior authentication is needed — a single crafted short frame from any device in radio range is sufficient. Frame validation in ieee802154_validate_frame() does not prevent it: a data frame is accepted with a one-byte payload.
The result is an out-of-bounds read and an out-of-bounds write of up to roughly 240 bytes into the adjacent network-buffer pool, corrupting other packets or allocator metadata and typically faulting the target. The out-of-bounds content is not attacker-chosen (it is ciphertext XOR keystream over out-of-bounds memory) and the frame is dropped when tag verification fails, so the primary impact is memory corruption and denial of service rather than information disclosure.
Exposure is limited to configurations that enable the experimental CONFIG_NET_L2_IEEE802154_SECURITY option, select a crypto device via CONFIG_NET_L2_IEEE802154_SECURITY_CRYPTO_DEV_NAME, and have established a security session with a level other than IEEE802154_SECURITY_LEVEL_NONE; with security disabled or at level NONE the tag length is zero and no underflow occurs. The fix rejects frames shorter than ll_hdr_len + authtag_len before the subtraction, and adds the matching guard on the transmit side in ieee802154_create_data_frame(). |
| The Bluetooth Mesh On-Demand Private Proxy solicitation handler in subsys/bluetooth/mesh/solicitation.c copies a received Solicitation PDU into a fixed 17-byte stack buffer without bounding the source length. In sol_pdu_decrypt(), out is allocated as NET_BUF_SIMPLE(17) and then filled with net_buf_simple_add_mem(out, in->data, in->len); net_buf_simple_add() guards its tailroom only with __ASSERT_NO_MSG, which is compiled out in production builds, so when in->len > 17 the underlying memcpy writes attacker-controlled bytes past the 17-byte stack buffer. The copy occurs before any decryption or authentication, so no key material is required to trigger it.
The oversized length arises because the mesh scan callback in subsys/bluetooth/mesh/adv.c calls net_buf_simple_restore() before dispatching to bt_mesh_sol_recv(), leaving buf->len covering the entire remaining advertising payload rather than just the Solicitation Service Data. After the parser locates the Service Data AD and consumes the Identification Type byte, the remaining buf->len is the 17-octet Network PDU plus any trailing advertising bytes, and prior to this fix there was no maximum-length check (only a minimum). An attacker can therefore append extra AD structures or padding after the Solicitation Service Data to make buf->len exceed 17.
bt_mesh_scan_cb() is registered directly as the BLE scan callback, so buf is raw, unauthenticated advertising data received over the air. Any device in radio range can send a non-connectable advertisement carrying a crafted mesh Proxy Solicitation to a node that has CONFIG_BT_MESH_OD_PRIV_PROXY_SRV enabled and is currently eligible to be solicited (GATT proxy disabled, On-Demand Private Proxy enabled), with no pairing, bonding, or provisioning. The result is an attacker-controlled stack overwrite — plausibly leading to remote code execution and at minimum a reliable remote denial of service. The fix trims buf->len to the spec-fixed 17 octets (dropping the PDU if fewer remain) before decryption. |
| The system-call verifier for i3c_do_ccc() in drivers/i3c/i3c_handlers.c validated the outer struct i3c_ccc_payload, the broadcast ccc.data buffer and the targets.payloads[] array, but did not validate the per-target data buffers those array elements point at. Each struct i3c_ccc_target_payload carries its own data pointer and data_len, and neither was passed through K_SYSCALL_MEMORY() before the payload was handed to z_impl_i3c_do_ccc() and on to the controller driver. The verifier also operated on the caller's live structure rather than a snapshot, so validated fields could be changed by a second user thread between the check and the driver's use — unlike the sibling z_vrfy_i3c_transfer(), which has always copied its message array first.
The defect is only present in CONFIG_USERSPACE builds, where drivers/i3c/i3c_handlers.c is compiled. An unprivileged user-mode thread that has been granted access to the I3C controller device object — the ordinary way an application lets a user thread talk to I3C peripherals — can issue a direct CCC whose target payload data pointer names an arbitrary kernel address. Controller drivers dereference that pointer directly (for example drivers/i3c/i3c_mcux.c, drivers/i3c/i3c_cdns.c, drivers/i3c/i3c_stm32.c, drivers/i3c/i3c_npcx.c), using rnw to decide direction.
A read CCC therefore causes the kernel-mode driver to write bus-received bytes into an attacker-chosen kernel address for an attacker-chosen length, and a write CCC transmits kernel memory out onto the I3C bus. The result is an out-of-bounds kernel write plus a kernel memory disclosure, i.e. escalation from a user-mode thread to supervisor privilege, defeating the isolation CONFIG_USERSPACE is meant to provide.
The fix introduces copy_ccc_and_do(), which snapshots the payload, copies the target array into kernel memory with k_usermode_alloc_from_copy() (bounding num_targets to fewer than 32), validates each per-target buffer with K_SYSCALL_MEMORY() according to rnw, and copies the driver-written num_xfer and err fields back to the caller. |
| The NXP GAU ADC driver (drivers/adc/adc_mcux_gau_adc.c) validated the caller-supplied sequence->buffer_size, which is expressed in bytes, against the number of active channels, which is a sample count. It then stored that byte count directly in data->results_length and used it in mcux_gau_adc_read_samples() as the number of uint16_t slots available. Because each conversion result occupies sizeof(uint16_t) bytes, a buffer that was accepted as "large enough" could be written with up to twice its size in bytes, so every sample past the buffer's midpoint was written out of bounds.
adc_read() and adc_read_async() are Zephyr system calls. The syscall verifier in drivers/adc/adc_handlers.c only confirms that the caller owns buffer_size writable bytes (K_SYSCALL_MEMORY_WRITE); deciding whether that size is sufficient for the requested channels and extra_samplings is delegated entirely to the driver. On a build with CONFIG_USERSPACE=y, a user-mode thread that has been granted the ADC device object could therefore submit a deliberately half-sized buffer and cause the driver's work-queue handler — which runs in supervisor mode, outside the caller's MPU restrictions — to write ADC conversion results past the end of that buffer, at an address and for a length of the caller's choosing.
The overrun is bounded by the requested sequence: with sequence->options->extra_samplings set, the sampling loop walks the buffer pointer forward across every sampling, so the total overrun can reach the full size of the supplied buffer (kilobytes for a large extra_samplings). The written words are 16-bit ADC conversion results, so the content is only partially attacker-influenced (via the selected analog input, gain and resolution), but the destination and length are fully controlled — sufficient for kernel memory corruption, a crash, or a userspace-to-kernel privilege escalation. Builds without CONFIG_USERSPACE, or on SoCs other than NXP RW61x with the GAU ADC node enabled, are not exposed to the privilege boundary; there the same defect only causes a silent overflow when the application itself passes an undersized buffer.
The fix replaces the ad-hoc check with the shared adc_sequence_validate_buffer() helper (validating against num_channels * sizeof(uint16_t)), stores buffer_size / sizeof(uint16_t) in results_length, and corrects the loop bound to a post-decrement so exactly the available number of slots may be written. |
