The Intel SEDI IPM (inter-processor mailbox) driver in drivers/ipm/ipm_sedi.c handles an inbound message interrupt in ipm_event_dispose(). It read the peer-written doorbell register, extracted the payload length with IPC_HEADER_GET_LENGTH(), and passed that length straight to sedi_ipc_read_msg() to copy the message into struct ipm_sedi_context.incoming_data_buf, without checking it against the buffer size. The doorbell length field is 10 bits wide (IPC_HEADER_LENGTH_MASK is 0x03FF), so it can encode up to 1023 bytes, while incoming_data_buf is IPC_DATA_LEN_MAX (128) bytes. The bounds check in the underlying HAL sedi_ipc_read_msg() is a DBG_CHECK that compiles away unless CONFIG_DEBUG is set, so no check remained in a production image. The doorbell register is written by the peer processor on the other side of the IPC link — for the intel_ish_5_* targets, the host CPU's ISH driver, reached through the device's memory-mapped register window. Host-side software with driver-level or raw BAR access can therefore set a length of up to 1023 and cause the interrupt handler to copy far past the destination buffer. The affected path requires an application to have registered an IPM receive callback via ipm_register_callback(), which is the driver's normal mode of use. The result is an out-of-bounds write of up to 895 bytes into static (.bss) memory, performed in interrupt context. The overflow first clobbers the rest of struct ipm_sedi_context — including the k_sem and k_mutex used by the transmit path, whose wait queues contain self-referential list pointers — and then adjacent static data, giving a kernel data-structure corruption and crash primitive. The overflowing bytes are read from registers following the message window, a portion of which are themselves peer-programmable. The fix rejects any doorbell whose encoded length exceeds IPC_DATA_LEN_MAX, logging it and acknowledging the doorbell so the peer is not left waiting.

The experimental USB host stack allocates a per-device configuration-descriptor buffer, udev->cfg_desc, from the dedicated usb_device_heap in usbh_device_set_configuration() (subsys/usb/host/usbh_device.c). On three failure paths — a failed full-length GET_DESCRIPTOR(CONFIGURATION) read, a mismatch between the short and full descriptor reads, and a rejected descriptor in parse_configuration_descriptor() — the buffer was released with k_heap_free() but the pointer was left dangling. The cleanup in usbh_device_free() is guarded only by if (udev->cfg_desc != NULL), so it frees the same block a second time. The path is driven entirely by the attached peripheral: usbh_device_connect() calls usbh_device_init(), which ends in usbh_device_set_configuration(), and on failure usbh_device_connect() calls usbh_device_free(). On v4.4.x this happens during the same enumeration, with no unplug required; on v4.1.0–v4.3.x the second free instead arrives via dev_removed_handler()/dev_connected_handler() in subsys/usb/host/usbh_core.c, so it requires a removal or duplicate-connect event after the failed enumeration — a sequence the attached device fully controls. A malicious or malformed USB device only has to answer the first 9-byte configuration-descriptor request with a well-formed header and then fail any of the three checks, for example by returning a full descriptor whose interface count disagrees with bNumInterfaces, or by answering the second read with different bytes. The result is a double free on usb_device_heap. On builds where lib/heap hardening is active (the current default CONFIG_SYS_HEAP_HARDENING_BASIC), sys_heap_free() detects the already-free chunk and calls k_panic(), giving a deterministic, peripheral-triggered denial of service of the USB host. On builds without that detection — earlier releases, or CONFIG_SYS_HEAP_HARDENING_NONE — the second free manipulates a chunk already on the free list, corrupting the heap's free list so that later allocations can return overlapping or invalid blocks. Exploitation beyond denial of service is bounded by the fact that usb_device_heap is a small dedicated heap (CONFIG_USBH_USB_DEVICE_HEAP, default 1024 bytes) whose only client is this descriptor buffer, and by CONFIG_USB_HOST_STACK being marked experimental and disabled by default. The fix sets udev->cfg_desc = NULL after every k_heap_free(), making the cleanup guard sound.

The default AEAD nonce provider for the PSA Internal Trusted Storage transform module, secure_storage_its_transform_aead_get_nonce() in subsys/secure_storage/src/its/transform/aead_get.c, stores its nonce counter in unsynchronized function-local static variables (s_nonce and s_nonce_initialized). Every ITS write obtains its AES-GCM or ChaCha20-Poly1305 nonce here via secure_storage_its_transform_to_store(). Because the function held no lock, two threads calling it concurrently race on the shared statics: the initialization path (psa_generate_random() followed by memcpy()) and the non-atomic increment-then-copy path can each hand the same nonce value to two distinct encryption operations, and can lose increments so the counter repeats values it was designed never to repeat. The ITS layer (secure_storage_its_set() in subsys/secure_storage/src/its/implementation.c) performs no serialization of its own, so concurrent same-UID writes reach the racy provider directly. Reusing a nonce with the same key under AES-GCM or ChaCha20-Poly1305 is a catastrophic AEAD failure: it leaks the XOR of the two plaintexts (ITS routinely stores secrets, including PSA persistent keys) and, for GCM, exposes the authentication key, enabling forgery of stored entries. Because the AEAD key is derived per entry UID, the security-relevant collision is two concurrent writes to the same UID both receiving the same nonce; an adversary able to read the raw backing storage can then exploit the reuse. Both ITS store back-ends shipped with Zephyr, zms.c and the settings/NVS back-end in settings.c, are log-structured flash stores with deferred garbage collection, so an entry superseded by a rewrite remains physically present in the partition until its sector is reclaimed. Two same-UID writes that race therefore leave both ciphertexts readable in the raw image at once, which is the condition the nonce reuse needs to be exploitable. The trigger remains narrow: both built-in key providers (DEVICE_ID_HASH and ENTRY_UID_HASH) salt the derived key with the entry UID, so reuse across different UIDs is harmless, and the exposure requires an application that writes the same UID concurrently from two threads. The fix serializes the provider with a K_MUTEX_DEFINE(s_nonce_mutex) held for the duration of nonce generation.

The Espressif ESP-hosted Wi-Fi driver (drivers/wifi/esp_hosted/) parses frames received over SPI from the ESP co-processor in esp_hosted_event_task(). For control frames it took the 16-bit TLV field data_length straight off the wire and passed it to pb_istream_from_buffer(frame.data_value, frame.data_length) without checking it against the frame length or the receive buffer. frame.data_value sits 26 bytes into a 3188-byte stack object, so a data_length of up to 0xFFFF makes pb_decode() read up to roughly 62 KB past the end of that object. Only the first fragment of a fragmented control response carries a TLV header; the pre-fix driver performed half-duplex SPI transactions and silently discarded any frame the co-processor queued while the host was transmitting (esp_hosted_hal_spi_transfer() aliased the RX buffer onto the TX buffer). When the discarded frame is the first fragment of a fragmented response, the driver treats the next fragment as a new frame — its per-fragment header and checksum are genuine, so both validation steps pass — and reads the TLV header out of raw protobuf continuation bytes. Those bytes come from control responses whose size and content an adjacent, unauthenticated attacker can influence, notably the AP scan list, which grows with the number and SSID length of access points in radio range. The impact is denial of service rather than disclosure. Reading past the end of the RAM region faults the device, and CONFIG_NANOPB_ENABLE_MALLOC is selected by the driver, so garbage length prefixes read out of bounds also drive heap allocations. The out-of-bounds bytes themselves do not reach the application: pb_decode() is started mid-stream on raw protobuf continuation bytes and so almost always fails outright, and anything that did decode would still have to pass esp_hosted_response(), which requires an exact msg_id match against the pending request, and then esp_hosted_ctrl_response(), which requires a success resp — an attacker influences the size and content of legitimate control responses, not the structure decoded out of misaligned bytes. Two related defects in the same receive path make the denial of service permanent: the fragment reassembly guard was sized with ESP_FRAME_SIZE instead of ESP_FRAME_MAX_PAYLOAD and, when tripped, returned from the sole RX thread instead of dropping the frame, and unhandled control events were queued with k_msgq_put(..., K_FOREVER) on an eight-entry queue that nothing drains, blocking that same thread. The driver has no watchdog or restart path, so either condition ends all Wi-Fi reception until the device is rebooted.

The Time-aware GPIO syscall verification handler z_vrfy_tgpio_pin_read_ts_ec() in drivers/timeaware_gpio/timeaware_gpio_handlers.c validated only the port device object and passed the caller-supplied timestamp and event_count output pointers to the driver without a K_SYSCALL_MEMORY_WRITE() check. The other handlers in the same file (z_vrfy_tgpio_port_get_time(), z_vrfy_tgpio_port_get_cycles_per_second()) already performed that check, so the omission left one syscall unguarded. tgpio_pin_read_ts_ec() is declared __syscall, so with CONFIG_USERSPACE=y an unprivileged user-mode thread that has been granted access to the TGPIO device object can invoke it with arbitrary pointer values. tgpio_intel_read_ts_ec() in drivers/timeaware_gpio/timeaware_gpio_intel.c bounds-checks only the pin index and then unconditionally performs timestamp = ... and event_count = ..., executing two 8-byte stores in supervisor mode at addresses chosen by the user-mode caller. The result is a write-what-where primitive that crosses the userspace/kernel boundary: the target address is fully attacker-chosen and the stored values are the hardware time-capture and event-counter register contents. Corrupting kernel data structures this way can escalate the calling thread to supervisor privilege or crash the system; the device-object permission required is a narrow capability that is not intended to confer any kernel-memory access. The fix adds the two missing K_SYSCALL_MEMORY_WRITE() validations before the driver call. Exposure is narrow in practice. Only builds with CONFIG_USERSPACE=y and CONFIG_TIMEAWARE_GPIO=y compile the affected file, and from v3.6.0 onward the file additionally referenced a relocated header (<zephyr/syscall_handler.h>) and removed Z_SYSCALL_* macros, so such a configuration failed to build until those were repaired after v4.4.0. Downstream trees that locally corrected that breakage, and v3.5.0 builds where it did not exist, are the exposed population.

gptp_handle_msg() in subsys/net/l2/ethernet/gptp/gptp.c dereferenced the gPTP header returned by GPTP_HDR() and switched on hdr->message_type without first checking that the received frame carries at least sizeof(struct gptp_hdr) (34) bytes of payload. The header accessor gptp_get_hdr() deliberately never fails for a short buffer — it returns pkt->frags->data and leaves validation to its callers — so a truncated frame produced a header pointer covering memory beyond the received data. The per-message-type checks that follow do not compensate: GPTP_VALID_LEN() reduces to len > 60 once the Ethernet header has been pulled, which is false for every fixed-size gPTP message, so GPTP_CHECK_LEN() never rejects a truncated SYNC, FOLLOWUP, PDELAY_RESP or SIGNALING message. The defect is reached by an unauthenticated peer on the same link sending an Ethernet frame with ethertype 0x88F7 to the PTP multicast address on an interface configured as a gPTP port, with CONFIG_NET_GPTP enabled. Because conformant Ethernet pads frames to 60 bytes, a payload shorter than 34 bytes generally requires a link that can deliver sub-minimum frames — for example the native_sim TAP driver (drivers/ethernet/eth_native_tap.c), which forwards whatever length the host device supplies, or a MAC configured to accept undersized frames. The short packet is retained (net_pkt_ref() into rcvd_sync_ptr, rcvd_follow_up_ptr, rcvd_pdelay_resp_ptr or rcvd_announce_ptr) and later parsed by the media-dependent and media-independent state machines in subsys/net/l2/ethernet/gptp/gptp_md.c and subsys/net/l2/ethernet/gptp/gptp_mi.c, which read tens of further bytes and copy some of them (the announce priority vector, hdr->port_id) into state that is subsequently transmitted. Under the default fixed-size buffer allocator (CONFIG_NET_BUF_FIXED_DATA_SIZE, 128-byte fragments) the accesses stay inside the allocated fragment and disclose stale recycled buffer contents; under the experimental CONFIG_NET_BUF_VARIABLE_DATA_SIZE allocator, where fragments are heap-allocated at the exact frame length, they are genuine out-of-bounds reads. There is no write and no availability impact.

net_icmpv6_send_error() in subsys/net/ip/icmpv6.c implemented only one of the three RFC 4443 section 2.4 suppression rules (do not answer an ICMPv6 error with an ICMPv6 error). It did not check whether the triggering packet's source address identifies a single node (rule e.6) or whether the packet was sent to a multicast destination (rule e.3, whose only exceptions are Packet Too Big and Parameter Problem Code 2). Of the five call sites, only the port-unreachable path in subsys/net/ip/connection.c carried an equivalent guard of its own; the extension-header, unknown-next-header and fragmentation paths in subsys/net/ip/ipv6.c and subsys/net/ip/ipv6_fragment.c had none. An unauthenticated attacker with access to the same link can exploit this in two ways. Sending a single IPv6 packet to the link-local all-nodes group ff02::1 carrying an unrecognized next-header value, with the source address spoofed to a chosen victim, causes every Zephyr node on the link to emit an ICMPv6 Parameter Problem message to that victim — a reflector with an amplification factor equal to the number of nodes. Alternatively, sending a unicast packet whose source address is a multicast address causes the node to transmit its ICMPv6 error to that multicast address, turning one unicast packet into a link-flooded multicast frame. Packets addressed to ff02::1 are accepted unconditionally by ipv6_input(), and no check rejects a multicast source address, so no special configuration is required. The impact is degraded availability of the shared link and of the reflection victim, together with the ability for the attacker to hide its own address behind the responding nodes. The effect is amplified on constrained mesh links such as 802.15.4/Thread, where link-local multicast is flooded hop by hop. There is no memory-safety consequence: the error packet itself is well formed, it is simply emitted in cases where the protocol forbids it. The fix adds both suppression checks at the single choke point in net_icmpv6_send_error(), before any reply packet is allocated, preserving the RFC-mandated exceptions for NET_ICMPV6_PACKET_TOO_BIG and Parameter Problem Code 2. Note that the IPv4 counterpart net_icmpv4_send_error() in subsys/net/ip/icmpv4.c still checks only for a broadcast destination and retains an equivalent gap for multicast destinations and non-unique sources.

gptp_mi_qualify_announce() in subsys/net/l2/ethernet/gptp/gptp_mi.c walks the Path Trace TLV of a received IEEE 802.1AS Announce message, comparing each clock identity against the local one. The loop bound was taken solely from the attacker-controlled wire field announce->steps_removed (accepted up to 254), never from announce->tlv.len, which is the field that states how many identities the TLV actually carries. Because path_sequence is the flexible member of the wire TLV (struct gptp_path_trace_tlv) and GPTP_ANNOUNCE() yields a raw pointer into the received packet buffer, the memcmp() inside the loop can address memory well past the end of the received frame. The stack's only length validation, GPTP_ANNOUNCE_CHECK_LEN(), requires the received gPTP payload to be exactly 68 + tlv.len bytes — so it does not constrain the loop, it guarantees the data is absent. An unauthenticated attacker on the same Ethernet segment can send a single Announce frame declaring tlv.len = 0 with steps_removed = 254; the frame passes the length check and reception path (net_gptp_recv() → gptp_handle_msg() → gptp_mi_qualify_announce()), which performs no authentication, and the loop then reads 255 entries of 8 bytes each — about 2 KB — beyond the end of the network buffer. The impact is an out-of-bounds read. The bytes read are only used as a memcmp() operand and are never returned to the attacker, so there is no meaningful information disclosure; the practical risk is that the overread crosses a network buffer pool boundary into unmapped or MPU-protected memory and faults the networking RX thread, causing a denial of service. Exposure is limited to builds that enable the opt-in, experimental CONFIG_NET_GPTP (TSN/AVB deployments) and to attackers with layer-2 adjacency, since gPTP frames are sent to a link-local multicast address and are not routed. The fix computes the true entry count as tlv.len / GPTP_CLOCK_ID_LEN and rejects the announce when steps_removed + 1 exceeds it, so the loop can no longer run past the data the packet-length check proved present.

Zephyr's TLS socket layer in subsys/net/lib/sockets/sockets_tls.c keeps a single process-global array, client_cache, of cached client sessions that is shared by every TLS socket context. The functions that mutate and read it — tls_session_save(), tls_session_get(), tls_session_cache_reset(), and the settings restore handler — allocate, free, and dereference each entry's heap buffer (entry->session). Before the fix these accesses were serialized only by the per-socket context mutex ctx->lock (assigned per socket in ctx_set_lock()), which provides no mutual exclusion between different sockets touching the shared cache. Because CONFIG_NET_SOCKETS_TLS_MAX_CLIENT_SESSION_COUNT defaults to 1, any two concurrent client sockets contend for the same slot. A thread in tls_session_get() reading entry->session inside mbedtls_ssl_session_load() can run concurrently with another thread in tls_session_save() that selects the same entry for reuse and executes mbedtls_free(entry->session) before reallocating — a use-after-free read, and a double-free when two saves evict the same entry. Both corrupt the mbedTLS heap. The cache is reached on ordinary client paths: at connect time via tls_session_store()/tls_session_restore(), and (on main) whenever a TLS 1.3 session ticket arrives during recv()/poll() via tls_session_store_current(). Exploitation requires an application that opts into per-socket client session caching (the TLS_SESSION_CACHE socket option, off by default) and runs concurrent TLS client connections on multiple threads; the timing that opens the window is influenced by the remote peer(s), so a malicious or compromised server can raise session-ticket frequency to widen it. The reliably-demonstrable impact is memory corruption leading to a crash or heap corruption (denial of service). The fix adds a dedicated session_cache_lock mutex taken across every accessor of client_cache, serializing all reads and frees and closing the race.

The ITE it82xx2 USB device-controller driver initialized its bus-suspend detection work with k_work_init_delayable(&priv->suspended_work, suspended_handler) inside it82xx2_enable() (the driver's .enable op) in drivers/usb/udc/udc_it82xx2.c. This work item is scheduled essentially continuously while the USB bus is active: the interrupt handler reschedules it on every SOF frame and suspended_handler() reschedules itself, so its timeout node is normally linked in the kernel timeout list / a workqueue pending queue. k_work_init_delayable() (kernel/work.c) unconditionally overwrites the entire k_work_delayable structure, including its timeout and queue linkage, with no busy check. Because it82xx2_disable() does not cancel the work, a normal disable-then-enable cycle re-runs api->enable() (udc_enable() only rejects a redundant enable, not a re-enable after disable) and re-initializes the still-pending work in place, corrupting the kernel timeout/workqueue linked lists and causing a kernel panic. An external USB host — for example a host performing USB DFU detach (dfu-util --detach) or forcing repeated attach/reset/re-enumeration — drives the udc_disable()/udc_enable() transitions and controls suspend/resume timing, so it can arrange for the suspend work to be pending across a re-enable. This yields an unauthenticated denial of service (kernel panic) reachable across the USB boundary from a removable, physically-connected host, with no confidentiality or integrity impact demonstrated. The fix moves the k_work_init_delayable() call into the one-time preinit function so the work is initialized exactly once, eliminating the re-initialization of an in-use item.