The Zephyr net_buf library (lib/net_buf/buf.c) manipulated both of its reference counts -- the per-header buf->ref and the per-data-block ref_count at the start of each variable/heap data allocation -- with plain non-atomic C operators (buf->ref++, if (--buf->ref > 0), if (--(*ref_count))). The API is documented as self-synchronizing: callers may share one buffer across threads (e.g. via k_fifo) and each holder independently calls net_buf_unref() with no surrounding lock. Under true concurrency (SMP, or single-core preemption between the non-atomic load and store while another context unrefs the same buffer), two holders can both observe the same prior reference value and both conclude they are the last reference. For heap/variable-data pools (mem_pool_data_unref/heap_data_unref, used by zbus message subscribers, the IP stack RX/TX buffers when CONFIG_NET_BUF_FIXED_DATA_SIZE=n, capture, wireguard, ISO-TP and usbip) this produces a double k_heap_free()/k_free() of the same block -- heap-metadata corruption and a use-after-free on the heap-hardening poison pattern. For the per-header refcount the buffer is returned to the pool free LIFO twice for any pool type (including fixed-data pools used by Bluetooth and networking), corrupting the free list so a later allocation hands the same buffer to two owners. The fix converts both refcounts to atomic_inc/atomic_dec (overlaying buf->ref in an atomic_t-sized union and changing the data-block refcount from uint8_t to atomic_t). Impact is gated on genuine concurrency and on an application architecture that shares one buffer among multiple independent unref'ers; the trigger is a refcount/timing race rather than packet content, so an external attacker has at most weak indirect influence over the race window. Affects all Zephyr releases through v4.4.0.

The Zephyr Bluetooth controller ISO Adaptation Layer (subsys/bluetooth/controller/ll_sw/isoal.c) fails to validate the length field of a framed ISO PDU start segment. Per the Bluetooth specification a start segment (sc=0) always carries a 3-byte time_offset, so its segment-header len must be at least PDU_ISO_SEG_TIMEOFFSET_SIZE (3). isoal_check_seg_header() accepted start segments with len < 3 as valid, and isoal_rx_framed_consume() then computed length = seg_hdr->len - 3 in a uint8_t, underflowing to 253-255 when len is 0-2. That oversized length is passed to isoal_rx_append_to_sdu(), whose copy is clamped only against the destination SDU buffer size, not the source PDU length, so up to ~255 bytes of controller memory beyond the received PDU are copied (via sink_sdu_write_hci()/net_buf_add_mem) into an HCI ISO data packet and delivered to the host. The PDU and its segment headers are entirely attacker-controlled and arrive over the air, reachable through both the CIS and BIS-sync HCI data paths (hci_driver.c) and the vendor data path (ull_iso.c), so a remote CIS peer or a broadcaster the device is synced to can trigger an out-of-bounds read causing information disclosure to the host and potential denial of service (faults or malformed oversized HCI ISO packets). The flaw affects all Zephyr releases since framed ISO reception was introduced in v3.0.0. The fix rejects sc=0 segments with len < 3 in isoal_check_seg_header() and adds a guard before the subtraction in isoal_rx_framed_consume().

A race condition in the Zephyr Bluetooth Classic RFCOMM host stack (subsys/bluetooth/host/classic/rfcomm.c) mishandles a simultaneous bidirectional session disconnect. When the local device has initiated a session teardown (state BT_RFCOMM_STATE_DISCONNECTING, DISC sent, RTX timer armed) and the connected peer concurrently sends its own DISC frame for dlci 0, rfcomm_handle_disc() invokes rfcomm_session_disconnected(), which unconditionally forced the session to BT_RFCOMM_STATE_DISCONNECTED without ever calling bt_l2cap_chan_disconnect(). Because the recovery timer was also cancelled and a later UA is ignored in the DISCONNECTED state, the session becomes permanently wedged: the underlying L2CAP channel is never released and the session slot in the fixed bt_rfcomm_pool[CONFIG_BT_MAX_CONN] array is never reclaimed (its conn pointer stays set). Subsequent bt_rfcomm_dlc_connect() calls on that connection fail with -EINVAL due to the invalid session state, so RFCOMM service is denied for that peer, and repeated occurrences can exhaust the session pool. The DISC frame is peer-controlled over the air, but exploitation requires the peer's DISC to collide with a local-initiated disconnect (a high-complexity timing race). Impact is availability/resource-leak only; there is no memory-safety, confidentiality, or integrity consequence. The defect shipped in released versions (present in v4.4.0 and earlier). The fix only transitions to DISCONNECTED when the session is not already in DISCONNECTING, preserving the proper L2CAP teardown path.

mcumgr_serial_process_frag() in subsys/mgmt/mcumgr/transport/src/serial_util.c calls net_buf_reset() on the result of smp_packet_alloc() before checking it for NULL. smp_packet_alloc() uses net_buf_alloc(K_NO_WAIT) against the shared MCUmgr packet pool (CONFIG_MCUMGR_TRANSPORT_NETBUF_COUNT, default 4), which returns NULL when the pool is exhausted. In default builds the __ASSERT_NO_MSG in net_buf_reset is a no-op, so net_buf_simple_reset writes through the NULL pointer (buf->len = 0; buf->data = buf->__buf), causing a fault/crash. The fragment data reaches this code from attacker-controlled bytes on the MCUmgr serial/UART/shell-console transports (smp_uart.c, smp_raw_uart.c, smp_shell.c), and a fresh buffer is allocated at the start of essentially every new packet. An attacker on the serial/console link can flood the transport to drive the 4-entry buffer pool to exhaustion and induce the NULL dereference, crashing the device (denial of service). The defect was introduced after the original MCUmgr rework and shipped in Zephyr v4.4.0. The fix moves the NULL check ahead of net_buf_reset.

The USB CDC-NCM device class (subsys/usb/device_next/class/usbd_cdc_ncm.c) ignores the return value of usbd_ep_enqueue() in its ethernet transmit callback cdc_ncm_send(). When the enqueue fails, the function still calls k_sem_take(&data-sync_sem, K_FOREVER), blocking on a completion semaphore that is only ever signaled from the bulk-IN transfer-completion callback. Because nothing was enqueued, that callback never fires and the calling thread — a shared network traffic-class TX thread — deadlocks permanently while holding the interface TX lock, halting transmission until reboot (and leaking the transmit buffer). The enqueue fails under conditions controlled by the attached USB host: usbd_ep_enqueue() returns -EPERM whenever the bus is suspended (a standard, persistent host operation), and the underlying udc_ep_enqueue() returns -EPERM/-ENODEV on disconnect, bus reset, or endpoint disable. The cdc_ncm_send() guard only checks the DATA_IFACE_ENABLED and IFACE_UP flags, not the suspended state, so a packet transmitted while the host holds the bus suspended reaches the failing enqueue and deadlocks the TX path. The realistic trigger is a bus suspend that occurs while the exported network interface is active and has traffic to send — host sleep, USB selective/auto-suspend, or hub power management — after which any device-originated packet deadlocks the path, recoverable only by reboot. The impact is a persistent loss of the virtual network connection between the host's NCM interface and the Zephyr device; because the deadlocked thread is a shared traffic-class TX thread, egress on other network interfaces can stall as well. There is no memory corruption or information disclosure. The defect was introduced with the CDC-NCM driver and shipped in releases through v4.4.0; it is fixed by checking the usbd_ep_enqueue() return value and freeing the buffer before the blocking wait.

Zephyr's HTTP server (subsys/net/lib/http) provides a static-filesystem resource type (HTTP_RESOURCE_TYPE_STATIC_FS, available when CONFIG_FILE_SYSTEM is enabled) that serves files from a configured root directory. Before this fix, both the HTTP/1 and HTTP/2 front-ends placed the raw, attacker-controlled request path into client-url_buffer (assembled in on_url() for HTTP/1 and copied verbatim from the :path pseudo-header for HTTP/2) without resolving ./.. segments. The static-FS handler then built the on-disk filename by directly concatenating the configured root with that raw URL (snprintk(fname, ..., "%s%s", static_fs_detail-fs_path, client-url_buffer) at http_server_http1.c:603 and http_server_http2.c:490) and opened it with fs_open(fname, FS_O_READ). Because the handler is reached via wildcard/leading-dir (fnmatch FNM_LEADING_DIR) or fallback resource matching, a request such as GET /<prefix/../../<file is dispatched to the handler and, after the underlying filesystem (e.g. LittleFS/FAT) resolves the .. segments, escapes the configured web root, letting an unauthenticated remote client read arbitrary readable files on the mounted volume (information disclosure). The HTTP server requires no TLS or authentication to reach this path. The fix adds http_server_remove_dot_segments(), which canonicalizes the path portion of the URL before resource lookup in both protocol handlers, neutralizing the traversal. Affects releases v4.0.0 through v4.4.0 for deployments that register a static-filesystem resource.

The IPv6 Neighbor Discovery handlers in subsys/net/ip/ipv6_nbr.c (handle_ra_input, handle_ns_input, handle_na_input) used an incorrect boolean expression that combined the RFC 4861 validity checks with the ICMPv6 code check using the wrong operator precedence: the form was '((length/hop/source/target checks) && (icmp_hdr-code != 0))'. Because every legitimate ND message carries ICMPv6 code 0, an attacker setting code == 0 (the normal value) caused the entire predicate to evaluate false, so the packet was never dropped and all of the other checks were silently skipped. The bypassed checks include the mandatory Hop Limit == 255 verification (which proves an ND packet originated on-link and was not forwarded) and, for Router Advertisements, the requirement that the source be a link-local address, as well as multicast-target sanity checks. As a result, an adjacent on-link attacker — and, because the Hop-Limit-255 guard is bypassed, potentially a remote/off-link attacker whose packets would otherwise be rejected — can have forged Router Advertisement, Neighbor Solicitation, and Neighbor Advertisement messages accepted. A forged RA lets the attacker reconfigure the victim's default router, on-link prefixes (SLAAC), MTU, reachable/retransmit timers, and (with CONFIG_NET_IPV6_RA_RDNSS) DNS servers, while forged NS/NA enable neighbor-cache poisoning, enabling man-in-the-middle, traffic redirection, and denial of service. The flaw is an input-validation/authentication weakness rather than a memory-safety issue: the underlying packet-parsing primitives (net_pkt_get_data, net_pkt_read, net_pkt_skip) are independently bounds-safe and the validated 'length' is the true buffer length, so skipping the length check causes no out-of-bounds access. The defect has existed since the logic was introduced in 2018 and shipped in all releases through v4.4.0; it is fixed by splitting the condition so any failing check drops the packet.

Zephyr's IP socket recvmsg() implementation (subsys/net/lib/sockets/sockets_inet.c, insert_pktinfo()) validated the user-supplied ancillary (msg_control) buffer using only the payload length (msg-msg_controllen < pktinfo_len) before writing a full control message consisting of an aligned cmsg header plus the payload. Because the check omitted the cmsg header size, a control buffer whose length falls in the under-checked window (e.g. 16-27 bytes for IPv4 IP_PKTINFO on a 64-bit target, where a single element actually occupies 28 bytes) passes the guard yet causes a fixed-size out-of-bounds write of up to one cmsg header (~12 bytes) past the end of the buffer. Under CONFIG_USERSPACE the recvmsg verifier allocates a kernel-heap copy of the control buffer sized to msg_controllen and runs the implementation against it, so the overflow corrupts kernel heap memory and is triggerable from an unprivileged userspace thread; in supervisor mode it corrupts the caller's buffer. The path is reachable on a UDP/IP socket with IP_PKTINFO/IPV6_RECVPKTINFO (or hoplimit/timestamping) enabled when the application calls recvmsg() with an undersized control buffer and a datagram is received; part of the overwritten bytes (the destination IP in ipi_addr) is influenced by the received packet. The fix makes the capacity check use NET_CMSG_SPACE(pktinfo_len) (aligned header + aligned data) and returns -ENOMEM when the buffer is too small. Affected: v3.6.0 through v4.4.0.

The Zephyr Bluetooth LE Audio Basic Audio Profile (BAP) unicast client mishandles peer-supplied ASE state notifications. In unicast_client_ep_qos_state() (subsys/bluetooth/audio/bap_unicast_client.c), the handler writes attacker-controlled QoS fields (interval, framing, phy, sdu, rtn, latency, pd) through the stream-qos pointer with only a stream != NULL guard. stream-qos is NULL for any stream that has been codec-configured via bt_bap_stream_config() but not yet added to a unicast group (it is set only by unicast_group_add_stream()). A malicious or buggy remote ASCS server, to which the local device is connected as a BAP unicast client, can send a GATT notification announcing the ASE has entered the QoS Configured state while the local endpoint is still in the Codec Configured state — a transition the dispatcher explicitly permits — during that window, causing a write through a NULL pointer and a crash (denial of service). The data written is itself remote-controlled. The defect shipped in v4.3.0 and v4.4.0 (and earlier). The fix re-points all BAP QoS storage to the always-valid embedded ep-qos struct, eliminating the NULL dereference.

Zephyr's BSD-sockets getaddrinfo() implementation (subsys/net/lib/sockets/getaddrinfo.c) passes a pointer to a stack-allocated state object (struct getaddrinfo_state ai_state) as the user_data of an asynchronous DNS resolver query. The socket layer waits on a semaphore with a timeout deliberately set slightly longer than the resolver's own per-query timeout. When that semaphore wait nonetheless times out (-EAGAIN) - which can occur when the resolver's timeout work is delayed by workqueue contention, or in the documented multi-retry configuration where CONFIG_NET_SOCKETS_DNS_TIMEOUT exceeds CONFIG_NET_SOCKETS_DNS_BACKOFF_INTERVAL - the pre-fix code retries the query (goto again) without cancelling the previous one and without resetting the semaphore. The previous query slot remains active in the resolver with its callback and the stack pointer as user_data, and ai_state-dns_id is overwritten so the stale query can no longer be cancelled. A subsequent DNS response delivered over UDP and matched by its 16-bit transaction id (in dispatcher_cb()/dns_read()), or the resolver's own delayed query-timeout work, then invokes dns_resolve_cb() against the now out-of-scope stack frame, writing through the stale pointer (state-status, state-idx, state-ai_arr[], and k_sem_give()). Because the triggering response is network-delivered and its 16-bit id is spoofable/replayable by an on- or off-path attacker, this is a network-influenceable use-after-return that can corrupt reused stack memory, leading to crashes/denial of service or memory corruption. The fix cancels the timed-out query by name and type before retrying and resets the local semaphore, eliminating the stale callback path. Affected: Zephyr v4.0.0 through v4.4.0.