79% vs 51%: independent neobanks lag on app security. Read the report

Deep Dive

Deep Dive

Linux KSMBD Pre-Auth Flaw, ZDI-Rated CVSS 9.0 RCE (ZDI-CAN-33774)

Linux KSMBD flaws are rated CVSS 9.0 by ZDI, one reachable before authentication. Linux 7.2.15 fixes the race; check vendor coverage and restrict SMB.

TL;DR

  • ZDI-CAN-33774: a pre-authentication race in KSMBD's protocol handler dispatch, triggered by one TCP write that pipelines a burst of SMB1 packets ahead of an SMB2 packet.
  • ZDI-CAN-33549: an integer underflow in KSMBD's SMB Direct (RDMA) path.
  • An unpatched host running KSMBD with TCP port 445 reachable by untrusted peers is exposed. ZDI-CAN-33774 needs no authentication. ZDI-CAN-33549 additionally needs the SMB Direct transport negotiated, an RDMA-capable interface and a writable share.
  • The public proof-of-concept code, posted with the reports on the public kernel lists, demonstrates a 28-byte kernel heap memory leak per attempt before authentication and a kernel crash, not code execution. ZDI rates both flaws as remote code execution and describes a write path for ZDI-CAN-33774, but nothing public shows it reaching code execution. No exploitation in the wild has been reported as of October 24, 2026 (Sysdig).
  • Linux 7.2.15 fixes ZDI-CAN-33774. ZDI-26-998, published November 13, 2026, adds the ZDI-CAN-33549 fix in Linux 7.3.2, 6.18.61 and 6.12.117, but not in 6.6.162 or 6.1.192. Neither flaw has a CVE id. Stop exposing KSMBD and leave SMB Direct off unless you need it.

Trend Micro's Zero Day Initiative rates both flaws CVSS 9.0 and credits Hyeontae Lee with both.

Am I affected?

ZDI-CAN-33774 exposes you to the demonstrated leak and crash if the first two conditions hold; ZDI-CAN-33549 additionally requires the third:

  • Your kernel build includes KSMBD and the service is running. systemctl status ksmbd reports the service state; lsmod | grep ksmbd reports whether the kernel module is loaded. A loaded module alone does not establish exposure: use the ss command below to check for a port 445 listener and confirm that KSMBD serves it.
  • TCP port 445 is reachable by the internet, tenant networks, or any other peer you do not fully trust.
  • For ZDI-CAN-33549 only: the connection can negotiate SMB Direct over an RDMA-capable interface, hardware or the rxe software driver, and a writable KSMBD share exists.

The two flaws differ in what they need and where they stand:

Flaw Where it lives What it needs Fix status as of November 16, 2026
ZDI-CAN-33774, CVSS 9.0 KSMBD protocol dispatch, fs/smb/server No authentication; any server running KSMBD on port 445 Backport aaf8a64a3fbf shipped in Linux 7.2.15, changelog published November 1, 2026; verify other kernels with their vendor
ZDI-CAN-33549, CVSS 9.0 KSMBD SMB Direct (RDMA) path SMB Direct negotiated, RDMA-capable interface, writable share The upstream fix commit 7193241a49c9 is in Linux 7.3.2, 6.18.61 and 6.12.117; the 6.6.162 and 6.1.192 branch tips still lack it; ZDI published ZDI-26-998 on November 13, 2026

Check what you run:

systemctl status ksmbd          # is the KSMBD service running?
lsmod | grep ksmbd              # is the module loaded?
ss -tlnp | grep :445            # is the SMB port listening, and on which address?

Where KSMBD actually runs:

Distribution or deployment What to check
Red Hat Enterprise Linux KSMBD is not included by default, per Red Hat.
AlmaLinux / CloudLinux 9 and 10 Their kernels do not build KSMBD, per CloudLinux.
Debian / Ubuntu KSMBD runs only after ksmbd-tools is installed and the service is started, per CloudLinux.
NAS / storage appliances Audit first: appliances that ship KSMBD-enabled kernels; check the running service and network exposure.

How do I fix it?

Linux 7.2.15 includes the ZDI-CAN-33774 race fix as backport aaf8a64a3fbf. Linux 7.3.2, 6.18.61 and 6.12.117 include the ZDI-CAN-33549 fix; the upstream commit is 7193241a49c9. Install a vendor-supported kernel that includes both fixes and reboot into it; check the vendor's advisory rather than assuming coverage from the version alone. Until coverage is confirmed, reduce exposure:

  1. Stop and disable KSMBD where you do not need kernel-space SMB serving: systemctl stop ksmbd && systemctl disable ksmbd. Trade-off: this removes the file service entirely.
  2. Where you must keep serving, restrict port 445 to networks whose hosts you trust, at the host firewall and any perimeter firewall. Trade-off: the service stays up, but only trusted peers can reach either flaw. Untrusted reach through any path leaves ZDI-CAN-33774 open, because it needs no authentication.
  3. Leave SMB Direct and RDMA off unless you specifically use them, which closes ZDI-CAN-33549. Trade-off: it does nothing for ZDI-CAN-33774, which is the pre-authentication flaw.

Check your distribution's kernel advisory for commit 68f410faedaf ("ksmbd: wait for negotiation before receiving another PDU") or its backport, aaf8a64a3fbf in Linux 7.2.15. A fixed source tree does not protect a host until it is running the corrected kernel. For ZDI-CAN-33549, check for commit 7193241a49c9, "smb: smbdirect: fix buffer length underflow in RDMA transfers", or its backport; the 6.6.162 and 6.1.192 branch tips did not contain it on November 16, 2026.

What is the root cause?

The KSMBD receive thread could read the next protocol data unit before the negotiation worker finished, so a pipelined SMB2 request could replace the connection's handler table while a worker was still handling an earlier, shorter SMB1 request.

The maintainer's own fix commit, 68f410faedaf, describes the mechanics exactly:

The receive thread queues a request and can read the next PDU before the negotiation worker finishes. While need_neg is set, the next PDU can cause ksmbd_init_smb_server() to replace conn->ops. A pipelined SMB2 request can therefore make the worker handle an earlier, shorter SMB1 request with SMB2 operations and read beyond its buffer.

An SMB1 multi-protocol negotiate may be followed by an SMB2 NEGOTIATE that changes conn->ops again.

The mechanics, step by step:

  • conn->ops is the connection's handler table: the set of function pointers that turn a decoded SMB command into an action. SMB1 and SMB2 have different command layouts, so the server assigns one table or the other during negotiation, when need_neg is set.
  • ksmbd_init_smb_server() is where that assignment happens. Re-running it mid-connection replaces conn->ops while work queued under the old table may still be running.
  • A pipelined request — several protocol data units packed into one TCP write — is what forces the overlap: the receive thread reads the SMB2 packet and re-runs negotiation before the worker that was queued for the earlier, shorter SMB1 packet has finished.
  • The worker then applies SMB2 handler logic to an SMB1-sized buffer and reads beyond its end.

For ZDI-CAN-33549, the report describes an integer underflow in the SMB Direct path: a client supplies RDMA buffer descriptors in which the first is longer than the declared payload, which makes the server's size counter wrap, so a subsequent memory copy overruns the end of an allocated buffer.

How is it exploited?

No exploitation in the wild has been reported, and the attack path is public in the reports themselves, posted to the public kernel mailing lists with proof-of-concept source code:

  1. The attacker connects to port 445 on a host running KSMBD. For ZDI-CAN-33774, no authentication is needed, which is why it exposes any server running KSMBD on that port.
  2. The attacker sends one TCP write containing a burst of short SMB1 packets followed by an SMB2 packet.
  3. The receive thread queues work for the SMB1 packets and reads the SMB2 packet, which re-runs negotiation and replaces conn->ops with the SMB2 handler table.
  4. The queued worker handles the earlier, shorter SMB1 request with SMB2 operations and reads past the end of the shorter buffer.
  5. What the public proof of concept demonstrates: 28 bytes of kernel heap memory leak per attempt, before authentication, and a kernel crash. ZDI's report also describes a path to writing past the buffer once an attacker has access to a share, but nothing public shows that path reaching code execution, Sysdig reported.

For ZDI-CAN-33549, the attacker is a client that negotiates SMB Direct and supplies crafted RDMA buffer descriptors; it needs an RDMA-capable interface on the server, a writable share, and the SMB Direct transport, so it sits behind more conditions than ZDI-CAN-33774.

How does the attack flow?

The whole attack is one pipelined TCP write that makes the server swap its handler table under a queued worker.

flowchart LR A["Attacker connects to
port 445, no authentication"] --> B["One TCP write: burst of
short SMB1 packets + SMB2"] B --> C["Receive thread reads SMB2
before the SMB1 worker finishes"] C --> D["Negotiation replaces
conn->ops with SMB2 handlers"] D --> E["Worker handles SMB1 data
with SMB2 handlers, reads past buffer"] E --> F["Kernel heap memory leaks
per attempt; kernel can crash"] style C fill:#f8d7da

The missing synchronization is at C: nothing makes the receive thread wait for the negotiation worker before it reads the next protocol data unit.

What can an attacker do?

Against an exposed host, the publicly demonstrated impact is kernel memory disclosure, 28 bytes of heap per attempt before authentication, and a kernel crash, which is a denial of service for a file server. ZDI rates both flaws CVSS 9.0 as remote code execution and its ZDI-CAN-33774 report describes a path to writing past the buffer once an attacker has share access, so memory corruption beyond the demonstrated leak is plausible. No public demonstration turns it into code execution, and no in-the-wild use has been reported.

The conditions each flaw needs:

  • ZDI-CAN-33774: to trigger the leak and crash, a kernel build with KSMBD, the service running, and port 445 reachable; no account or share. ZDI's described write path additionally requires share access.
  • ZDI-CAN-33549: all of the above plus SMB Direct negotiated, an RDMA-capable interface (hardware or the rxe software driver) and a writable share.
  • Neither flaw affects userspace Samba (smbd), which is a separate codebase that does not share KSMBD's kernel dispatch path.

Why does the fix work?

The fix for ZDI-CAN-33774, commit 68f410faedaf by KSMBD maintainer Namjae Jeon, adds a wait to the receive path in fs/smb/server/server.c:

--- a/fs/smb/server/server.c
+++ b/fs/smb/server/server.c
@@ static int queue_ksmbd_work(struct ksmbd_conn *conn)
    struct ksmbd_work *work;
+   bool wait_for_neg = ksmbd_conn_new(conn) || ksmbd_conn_need_negotiate(conn);
    int err;
@@
    INIT_WORK(&work->work, handle_ksmbd_work);
    ksmbd_queue_work(work);
+   /* Negotiation can change conn->ops, finish it before reading another PDU. */
+   if (wait_for_neg)
+       wait_event(conn->req_running_q,
+              atomic_read(&conn->req_running) == 0);
    return 0;

After queuing work on a connection that is new or still negotiating, the receive thread now waits on conn->req_running_q until every queued request has finished. The handler table cannot be replaced while a worker is still handling an earlier request, so SMB1 data can never be handled with SMB2 operations. The commit keeps need_neg semantics unchanged. Its message frames the change as keeping the operations stable for both stages of negotiation.

The fix addresses this race and credits ihopenre-eng for a related earlier report. Linux 7.2.15 includes it, but its changelog alone does not establish coverage in another stable branch or a distribution kernel. It does not change the SMB Direct path, so ZDI-CAN-33549 requires a separate fix, commit 7193241a49c9.

How do I check if I was compromised?

No exploitation has been reported and no indicators of compromise have been published for either flaw. If you run KSMBD exposed to untrusted networks and want to audit:

  • Review port 445 connection and firewall logs for unexpected source addresses. Where packet captures are available, look for bursts of requests followed by abrupt disconnects; this pattern alone does not establish exploitation.
  • Check kernel logs for oops or panic records referencing ksmbd or fs/smb/server, since the public proof of concept can crash the kernel.
  • Treat the 28-byte-per-attempt heap leak as a real exposure: each attempt discloses kernel heap content, so repeated attempts against an exposed host may have harvested memory contents even without a crash. If compromise is suspected on a host, the standard response for a suspected kernel-level incident applies: isolate the host, rotate every credential it holds, and rebuild from a known-good image.

Timeline

  • 2026-09-30: KSMBD maintainer Namjae Jeon writes fix commit 68f410faedaf, "ksmbd: wait for negotiation before receiving another PDU", crediting ihopenre-eng, responding to an earlier report.
  • 2026-10-20: the fix commit merges into the mainline Linux development tree (committer timestamp October 20, 23:55 UTC).
  • 2026-10-21: ZDI discloses ZDI-CAN-33774 and ZDI-CAN-33549, each rated CVSS 9.0 and credited to Hyeontae Lee, to the public kernel mailing lists with no patches attached (UTC timestamps).
  • 2026-10-21: SMB Direct maintainer Stefan Metzmacher asks the reporter to test the for-7.5/smbdirect development branch as a possible fix for ZDI-CAN-33549.
  • 2026-10-22: KSMBD maintainer Namjae Jeon replies that commit 68f410faedaf already fixes the race.
  • 2026-10-22: stable-kernel maintainer Greg Kroah-Hartman replies asking for patches rather than public disclosures.
  • 2026-10-22: The Hacker News covers the disclosures.
  • 2026-10-23: Namjae Jeon posts a smaller SMB Direct underflow patch after Stefan Metzmacher recommends a minimal fix.
  • 2026-10-23 to 25: AlmaLinux, CloudLinux and Sysdig publish analyses; the public proof of concept is documented as a 28-byte kernel heap leak per attempt before authentication plus a crash, not code execution.
  • 2026-10-26: the ZDI-CAN-33549 fix commit 7193241a49c9, "smb: smbdirect: fix buffer length underflow in RDMA transfers", merges into the mainline Linux development tree.
  • 2026-10-26: this post. No exploitation in the wild has been reported, and neither flaw is in the CISA KEV catalog.
  • 2026-11-01: Linux 7.2.15 ships the ZDI-CAN-33774 fix, as backport commit aaf8a64a3fbf, per kernel.org's changelog.
  • 2026-11-13: ZDI publishes ZDI-26-998 for ZDI-CAN-33549 with fix commit 7193241a49c9.

FAQ

Why do these flaws have no CVE id?

The cited reports use ZDI-CAN identifiers and do not provide a CVE id for either flaw. The Linux kernel policy says automatic CVE assignment follows a fix applied to a stable kernel tree; it also allows requesting a reserved identifier before a fix. Linux 7.2.15 includes the ZDI-CAN-33774 fix, but that release alone does not establish CVE assignment. Track the original fix commit as well as any later CVE announcement.

Does the public proof of concept give attackers remote code execution?

No. What it demonstrates is a 28-byte kernel heap memory leak per attempt before authentication and a kernel crash. ZDI's CVSS 9.0 ratings treat the flaws as remote code execution, and its ZDI-CAN-33774 report describes a path to writing past the buffer once an attacker has share access, but no public demonstration shows code execution, Sysdig reported.

Is my distribution's kernel affected?

Only if the distribution builds KSMBD, the service has been started and the running kernel lacks the fix for the relevant flaw. Red Hat Enterprise Linux does not include KSMBD by default, AlmaLinux and CloudLinux 9 and 10 kernels do not build it, and Debian and Ubuntu run it only after ksmbd-tools is installed and the service started. Hosts to audit are NAS appliances and storage servers that ship kernels with KSMBD enabled to serve files.

Are Samba servers affected?

No. Samba's userspace smbd is a separate implementation that does not share KSMBD's kernel dispatch code. The exposed software is KSMBD, the in-kernel SMB server, identifiable with systemctl status ksmbd or lsmod | grep ksmbd.

When will a fix reach my kernel?

The ZDI-CAN-33774 fix reached the mainline development tree on October 20, 2026, and Linux 7.2.15 shipped it on November 1, 2026, per kernel.org's changelog. The ZDI-CAN-33549 fix was committed on October 26, 2026 as 7193241a49c9 and shipped in Linux 7.3.2, 6.18.61 and 6.12.117; the 6.6.162 and 6.1.192 branch tips still lack it. Check other stable lines and distribution kernels separately, and confirm coverage of each flaw with your vendor.

Do I need SMB Direct to be worried about both flaws?

No. ZDI-CAN-33549 needs SMB Direct, an RDMA-capable interface and a writable share. ZDI-CAN-33774 needs none of that: it is reachable by any unauthenticated peer that can reach port 445 on a host running KSMBD, which is why it is the one to remove exposure to first.

Find exposed SMB servers before an attacker does

Follow Ostorlab's network scan guide: select New Scan → Network, enter your IP addresses or ranges, choose Full Network Scan, and start the scan. Review any port 445 findings, then confirm locally whether the service is KSMBD or userspace Samba. Restrict exposed KSMBD services while waiting for a fixed distribution kernel.

References

  • Primary ZDI-CAN-33774 report, mirrored by Openwall, with the October 21 UTC disclosure timestamp and described write path: https://lists.openwall.net/linux-kernel/2026/10/21/2694
  • Stefan Metzmacher's SMB Direct branch and minimal-fix discussion, October 23: https://lists.openwall.net/linux-kernel/2026/10/23/739
  • Namjae Jeon's attached SMB Direct underflow patch, October 23: https://lists.openwall.net/linux-kernel/2026/10/23/971
  • The Hacker News, "Two CVSS 9.0 Linux KSMBD RCE Flaws Disclosed on Public Mailing Lists Before Kernel Fixes", October 22, 2026, updated October 26, with the flaw mechanics, the maintainer replies, the proof-of-concept scope and the distribution status: https://thehackernews.com/2026/10/two-cvss-90-linux-ksmbd-rce-flaws.html (checked October 26, 2026)
  • ZDI-26-998, the published advisory for ZDI-CAN-33549, CVSS 9.0, vendor Linux, discovered by Hyeontae Lee, with the fix commit: https://www.zerodayinitiative.com/advisories/ZDI-26-998/ (checked November 16, 2026)
  • Fix commit 7193241a49c9, "smb: smbdirect: fix buffer length underflow in RDMA transfers", authored October 23, 2026, committed October 26, 2026: https://github.com/torvalds/linux/commit/7193241a49c958782ba4b8df06da3734d1e317d2 (checked November 16, 2026)
  • The ZDI report for ZDI-CAN-33774 on the public kernel lists: https://lore.kernel.org/lkml/BYAPR01MB383076AC412E19C48CFBEAE980662@BYAPR01MB3830.prod.exchangelabs.com/ (primary report)
  • The ZDI report for ZDI-CAN-33549 on the public kernel lists: https://lore.kernel.org/lkml/BYAPR01MB3830472B5702C089B0BBB77D80662@BYAPR01MB3830.prod.exchangelabs.com/ (primary report)
  • Fix commit 68f410faedaf, "ksmbd: wait for negotiation before receiving another PDU", in the mainline Linux tree, authored September 30, 2026, committer timestamp October 20, 2026: https://github.com/torvalds/linux/commit/68f410faedaf6ee74d81e574733c50ebcc1bd8ed (checked October 26, 2026)
  • Sysdig on the six October kernel reports, the proof-of-concept scope and the unmerged fix status: https://www.sysdig.com/blog/detecting-and-mitigating-zdi-can-33367-and-five-other-linux-flaws (referenced via The Hacker News, October 26, 2026)
  • AlmaLinux on the public kernel-list reports: https://almalinux.org/blog/2026-10-23-zdi-can-33367/ (referenced via The Hacker News, October 26, 2026)
  • CloudLinux on the KSMBD proof of concept and distribution build status: https://blog.cloudlinux.com/zdi-can-33367 (referenced via The Hacker News, October 26, 2026)
  • Red Hat solution document 6991749, KSMBD not included in RHEL by default: https://access.redhat.com/solutions/6991749 (referenced via The Hacker News, October 26, 2026)
  • The Linux kernel CVE assignment policy, CVEs issued only after a stable-tree fix: https://docs.kernel.org/process/cve.html (referenced via The Hacker News, October 26, 2026)
  • Linux 7.2.15 changelog, naming backport aaf8a64a3fbf of the ZDI-CAN-33774 fix: https://cdn.kernel.org/pub/linux/kernel/v7.x/ChangeLog-7.2.15 (checked November 1, 2026)