Skip to content

CISA KEV September 18, 2026

Tags

Summary

CISA's September 18, 2026 KEV additions (catalog 2026.09.18, 1,716 entries, a third row landed later in the day, raising the batch from two to three) are three Linux kernel vulnerabilities, ALL THREE with CISA SSVC assessments published Sep 18, 2026 grading exploitation "active," all on a compressed BOD 26-04 due date of 2026-09-21, three days, and Forensics Triage requirements on all three. None names an actor; ransomware use is unknown on all three. Two 2025-vintage CVEs receiving 72-hour KEV deadlines ~12 months after their fixes shipped, and a June-2026 CVE getting one three months after its fix, is itself the exploitation-signal event.

CVE Subsystem Class CVSS (kernel CNA) CWE KEV added BOD due CISA SSVC (Sep 18)
CVE-2025-39682 net/tls (kTLS RX path) Improper check for unusual/exceptional conditions, a zero-length record retrieved from the rx_list bypasses the intended recvmsg() record-type handling, so subsequent TLS records are processed under incorrect zero-copy and queuing assumptions 9.8 CRITICAL (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H), remote, no auth, no user interaction CWE-754 2026-09-18 2026-09-21 Exploitation: ACTIVE · Automatable: YES · Technical Impact: TOTAL
CVE-2025-39964 crypto: af_alg Race condition, concurrent writes to the same AF_ALG socket interleave data unpredictably and corrupt internal socket state 7.8 (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H) CWE-362 2026-09-18 2026-09-21 Exploitation: ACTIVE · Automatable: NO · Technical Impact: TOTAL
CVE-2026-53266 netfilter: bridge (ebtables SNAT target) Out-of-bounds write, the ARP sender-hardware-address rewrite writes directly into a nonlinear socket-buffer fragment backed by a splice-imported file page 8.8 (AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H, scope-changed) CWE-787 2026-09-18 2026-09-21 Exploitation: ACTIVE · Automatable: NO · Technical Impact: TOTAL

All three rows carry the standard kernel language, "could be end-of-life (EoL) and/or end-of-service (EoS)… discontinue use and/or transition to a supported version", and BOD 26-04 risk-based patching plus forensic-triage requirements. Note the CVSS split on CVE-2025-39682: the kernel CNA scored it 9.8 fully remote; NVD's own primary assessment scored it 7.1 local (AV:L/PR:L), see the CVE section below for why the CNA's remote reading is the one that matches the commit.

CVE-2025-39682, kTLS zero-length record on the rx_list (the remote one)

  • Mechanism (from the fix, tls: fix handling of zero-length records on the rx_list): kTLS recvmsg() must process either contiguous DATA records or exactly one non-DATA record per call; when a record has already been decrypted (TLS 1.3, record type is unknown until decryption) it is queued to rx_list for the next recvmsg(). Zero-copy decryption (decrypted straight into the user buffer, no skb to queue) is only allowed for DATA records, so the type-change-after-zero-copy case was assumed impossible, except when the initial record picked up from rx_list is zero-length, the corner case the code missed. A subsequent record is then processed under the wrong zero-copy/queuing assumptions: kernel memory-corruption mechanics in the TLS receive path.
  • Why the CNA calls it remote: the kernel CNA's CVSS scenario is explicit, "the bug is triggered entirely by the sequence and content of TLS records arriving from the remote peer on a kTLS-enabled TCP socket; no local access is needed," and AC:L because three back-to-back records ([DATA][zero-length non-DATA][DATA]) make the corruption deterministic, no race to win, no memory-layout dependency (a kernel selftest ships with the fix). Any kTLS RX consumer is remotely reachable: OpenSSL-with-kTLS servers/clients, NFS-over-TLS, SMB/RPC-over-TLS via net/handshake (net/tls/tls_sw.c is the affected file). Treat this as remote unauthenticated kernel memory corruption, not a postfoothold bug, NVD's 7.1 AV:L primary appears to reflect a local-trigger interpretation; the SSVC adds Automatable: YES (the only one of the three), the combination CISA reserves for worm-adjacent risk.
  • Same late-listing story as its batch-mates: CVE published 2025-09-05, fix upstream Aug 2025 with stable backports cited in the KEV row, then ~12 months un-KEV'd before a 72-hour deadline. CISA does not state what changed; the listing plus SSVC "active" is the signal.
  • Public exploit code has existed since October 2025: a published kernelCTF 1-day exploit for this CVE (a variant of the CVE-2024-58239 zero-length-record family) has been on GitHub since 2025-10-14, attacker-reachable working code against the same mitigation-* kernel builds Google's kernelCTF runs, so the "what changed in the exploitation picture" question already has a public answer for anyone who looked: weaponization was not the missing piece; KEV listing lagged the public 1-day by a year.
  • Durable defense: the exposed surface is hosts with kTLS offload enabled (kernel-config CONFIG_TLS, plus any consumer that negotiates it). Identify TLS terminators relying on kernel TLS rather than userspace TLS; patch is the answer, there is no practical workaround short of disabling CONFIG_TLS/kTLS usage, and SSVC "active + automatable" makes the 3-day federal deadline the right tempo for internet-facing kTLS hosts.

CVE-2025-39964, AF_ALG concurrent-write race

  • Upstream fix: crypto: af_alg - Disallow concurrent writes in af_alg_sendmsg, the commit message is blunt: "Issuing two writes to the same af_alg socket is bogus as the data will be interleaved in an unpredictable fashion. Furthermore, concurrent writes may create inconsistencies in the internal socket state." Fixed by adding a ctx->write ownership field (multiple stable-tree commits, Sep-Oct 2025, cited in the KEV row).
  • Timeline is the story: this CVE was published 2025-10-13 and sat un-KEV'd for ~11 months before CISA listed it on a 3-day deadline. KEV additions long after a fix ships are how CISA signals the exploitation picture changed, either in-the-wild use surfaced or field evidence accumulated. CISA does not state which; treat the listing as the signal, not a confirmed campaign.
  • Why AF_ALG matters: AF_ALG is the kernel's socket-based userspace crypto API, reachable from unprivileged processes (historically via unprivileged user namespaces). It has a deep history as an LPE primitive family precisely because the internal af_alg state machine handles async I/O with shared buffers. The fix disallows concurrent writes outright rather than serializing them, a hint at how untrustworthy the interleaved path was.
  • Durable defense: AF_ALG is not needed on most hosts, blocklisted by many hardened kernels and container runtimes already. modprobe blacklist algif_aead/algif_skcipher (or disable AF_ALG socket creation via seccomp in container profiles) removes the class regardless of patch state.

CVE-2026-53266, ebtables SNAT rewrite into splice-shared file pages

  • Upstream fix: netfilter: bridge: make ebt_snat ARP rewrite writable, at the bridge ebtables hooks the Ethernet header is addressed through skb_mac_header()/eth_hdr(), while skb->data points at the payload; the SNAT target's ARP sender-hardware-address rewrite was not preceded by a writability/linearization check for the nonlinear case, so when the header region sat in a nonlinear skb fragment backed by a page imported via splice (i.e., a page-cache file page shared with the original file), the rewrite modified shared file-backed memory directly, an out-of-bounds/shared-memory write class of the Dirty Pipe/dirty pagetable family, where corrupting page cache contents is the LPE lever. Scope-changed CVSS 8.8 (kernel CNA), published 2026-06-25, stable commits cited in the KEV row.
  • Preconditions matter for triage: the vector is local (AV:L, low privileges) and reaches the bridge ebtables SNAT path with CAP_NET_ADMIN-adjacent namespace access, the realistic ITW shape is a container or unprivileged-user-namespace foothold using ARP-path crafted frames against bridge processing. That is exactly the shared-hosting / container-host kill chain: webshell → namespace foothold → kernel LPE → root on every tenant's workload.
  • Durable defense: the class fix is systemic, reject nonlinear skb writes to shared pages; the durable host control is that unprivileged user namespaces + network namespace creation = kernel attack surface, and hosts running untrusted code (shared hosting, CI runners, agent sandboxes) should gate net/bridge/ebtables capabilities in the namespace, or disable unprivileged userns entirely where the workload allows.

Defender takeaways

  • Read the batch through SSVC, not just the CVSS column: CISA graded ALL THREE "active exploitation" on Sep 18, and only the kTLS bug "Automatable: yes." A remote, deterministic, no-race kernel TLS-corruption bug with active exploitation is the loudest signal in the set, triage kTLS-enabled internet-facing hosts first, then the two local escalations.
  • Three-day BOD deadline on all three (2026-09-21) is the urgency signal. Federal assets must patch or mitigate in 72 hours; that cadence is what CISA reserves for exploitation-informed items.
  • Two 2025-vintage kernel CVEs listed ~12 months after their fixes shipped, plus a fresh June-2026 CVE on the same 72-hour window, all SSVC-active the same day says CISA holds field evidence across the kernel exploit space, not one campaign. Treat the listings jointly: something changed in the exploitation picture for kernel primitives broadly.
  • The AF_ALG and ebtables bugs are postfoothold LPEs; the kTLS bug is not. The defensive question for the LPE pair is what your local-foothold blast radius looks like: a kernel LPE converts a single webshell or container escape into whole-host compromise, and on shared hosts every tenant with it. The kTLS bug reaches you before any foothold.
  • Patch level: apply the stable-tree commits per the KEV notes (kernel.org links in each row) or the distro kernel errata that carry them; EoL kernels (the KEV's explicit EoL/EoS note) have no fix path, migrate or remove from internet/namespace-reachable hosts.
  • Forensics Triage = Yes on all three: preserve memory/disk evidence per the BOD 26-04 triage guidance before re-imaging; a page-cache-corruption LPE leaves file-backed evidence in odd places, and TLS-path memory corruption may leave evidence only in memory.
  • Container-runtime hardening closes the two LPE classes ahead of patch cycles: seccomp profiles that deny socket(AF_ALG) and drop CAP_NET_ADMIN in user namespaces remove the reach for both bugs; the CVEs are then depth, not the primary control. No comparable mitigation exists for kTLS short of patching, inventory which TLS terminators use kernel TLS.
  • Pattern read: this is the second September 2026 KEV group arriving on compressed deadlines with the EoL warning language (after the Cisco ISE batch on 3-day-comparable windows), CISA's 2026 cadence increasingly treats "supported version" as the remediation, not "patch the version you're on."

September 19 follow-up: a same-week public exploit wave for FOUR more kernel LPEs (DirtyAH6 / TUNderflow / PPPoEject / DiagSpill)

The day after this batch was KEV-listed, researcher Asim Manizada published working exploit code for four additional local-root kernel flaws (primary write-up "A quartet of Linux local root vulns," heyitsas.im, Sep 18 06:00 UTC per the linux-distros embargo; The Hacker News same day): DirtyAH6 (CVE-2026-80844, IPsec AH6/IPv6), TUNderflow (CVE-2026-81000, TUN/TAP), PPPoEject (CVE-2026-68121, PPPoE), DiagSpill (CVE-2026-74469, SCTP sctp_diag). All four are fixed upstream, first stable releases carrying the complete set: 5.10.270 / 5.15.221 / 6.1.188 / 6.6.157 / 6.12.109 / 6.18.50 / 7.2.4 (match distro errata, not these numbers); reported to security@kernel.org mid-July; no ITW use reported; exploits are build-tuned and crash-prone. Underlying bugs are 10-21 years old.

Verified primary-source mechanics worth keeping: DirtyAH6, ipv6_rearrange_rthdr() moves the address pointer by segments - segments_left without checking segments_left ≤ segments; a raw IPv6 HDRINCL packet with hdrlen=2, segments_left=255 drives a 4,064-byte back-memmove out-of-bounds; on an IPv6 router/gateway that adds AH in transport mode the same bug is a remote crash, and the researcher reached remote root in-lab only with on-target memory grooming (remote-only grooming "extremely difficult," not ruled out). TUNderflow, tun_set_headroom() stores receive headroom in tun->align; a netkit device under VXLAN-under-OVS can push 4,160 bytes to a raw TUN port so SKB_MAX_HEAD(4160) underflows and tun_alloc_skb() leaves skb->data 64 bytes past its allocation. PPPoEject, pppoe_sendmsg() keeps a pointer into the skb head across dev_hard_header(), where a lower-device callback can pskb_expand_head() and free it (triggered by FUSE-blocking the payload copy while adding the first GRE port to an empty team/bond) → stale-pointer writes. DiagSpill, SCTP allows 65,536 peer transports but transport_count is 16-bit, so the 65,536th wraps to 0 and sctp_diag copies the full list with no reserved payload, spilling ~8 MiB past the Netlink response; remotely triggerable crash only if ASCONF/ADD-IP + SCTP-AUTH (or addip_noauth_enable=1) are enabled, all off by default, and the author sees no path to remote root for this one. Notably, the researcher credits discovery to LLM-assisted harness tooling ("Getting LLMs Drunk to Find Remote Linux Kernel OOB Writes" lineage, graph-based object tracking + 'geometric' memory-state agents), the same frontier-AI vuln-discovery pattern as Unit 42 NOVA on-wiki, now producing embargoed kernel LPEs.

The connection to this page is the precondition overlap: three of the four require unprivileged user namespaces, and DiagSpill requires no privilege at all, only the SCTP module, exactly the userns/netns surface this page's class-controls section tells shared hosts and CI/runner operators to gate. Read jointly with the KEV batch: CISA's Sep 18 "active across kernel exploit space broadly" grading and a same-week publication of four fresh LPE exploits mean the public weaponization pipeline for kernel LPEs is running regardless of any single CVE's KEV status. The same class controls (disable/gate unprivileged userns, deny socket(AF_ALG), drop CAP_NET_ADMIN in userns, blacklist sctp where unused) close all four ahead of errata. None of the four is KEV-listed as of this wiki's Sep 19-20 catalog check (2026.09.18, 1,716 entries, no bump). Monitor: KEV additions for any of the four, distro-errata rollout, copycat publication against the KEV trio (this page's standing watch item).

September 21 post-deadline state: the BOD passed with no catalog change, no attribution, and no fourth row

The compressed 3-day BOD 26-04 due date for all three kernel rows was 2026-09-21. At this wiki's post-deadline check (Sep 21 ~09:25 UTC, earlier same-day checks at 05:26 and 07:25 UTC were identical), the KEV catalog is UNCHANGED: version 2026.09.18, 1,716 entries. Since listing: no catalog bump (no fourth quartet row, DirtyAH6 / TUNderflow / PPPoEject / DiagSpill, all with public exploit code since Sep 18-19, remain UNLISTED), no actor attribution appended to any row, no SSVC re-grading visible in the JSON, and no CISA/FBI joint advisory tying the three same-day listings to a named campaign. The newest catalog entries are still the Sep 18 kernel trio; the previous batch rows (Pixel CVE-2026-58704, Cisco ISE CVE-2026-76460, Acronis CVE-2026-87886 of Sep 16) sit below them unchanged. Monitor: whether a Sep 22 batch absorbs the public-exploit quartet (the Sep 18 precedent says public exploit code + kernel LPE is exactly what CISA adds), whether any distro errata cycle finally names an exploitation vector, and whether the kTLS row (SSVC active + automatable, the only automatable item of the month) ever gets an ITW report beyond the standing kernelCTF 1-day. Standing null result to preserve: three SSVC-"active" kernel CVEs went through their entire 72-hour federal deadline with zero public corroboration of active exploitation, either CISA's field evidence stays classified-level quiet, or the "active" grading is doing more work than the word usually carries; either way, this page's class controls (gate unprivileged userns, deny socket(AF_ALG), drop CAP_NET_ADMIN in userns, inventory kTLS consumers) remain the right posture regardless of what attribution never arrived.

September 21 (evening) catalog bump: one new row, Zyxel GS1900 series switches, while the kernel trio still sits un-adorned

At the eighth-sweep check (Sep 21 ~21:00 UTC) the catalog moved for the first time since Sep 18: version 2026.09.21, 1,717 entries. The single new row is CVE-2026-7273, Zyxel GS1900 series switches, stack-based buffer overflow (CWE-121) in the CGI program; LAN-based unauthenticated attacker → OS command execution via crafted HTTP request; added Sep 21, due 2026-09-24, Forensics Triage = Yes, ransomware use Unknown, requiredAction cites BOD 26-04 + Forensics Triage (vendor advisory dated 2026-06-16, i.e., four months between Zyxel's advisory and the KEV listing, a reminder that KEV arrival is not the patch clock, it's the exploitation clock). No actor named.

What did NOT change: the three kernel rows carry no attribution annotation, the public-exploit quartet (DirtyAH6 / TUNderflow / PPPoEject / DiagSpill) is still unlisted, and no joint advisory tied the Sep 18 trio to a campaign. The post-deadline null result recorded above stands, with one refinement: CISA clearly kept cataloging after the deadline (a Zyxel row landed the same evening), the silence is specific to the kernel trio, not institutional. Monitor: whether a Sep 22 batch absorbs the quartet, and whether the Zyxel row (LAN-based access = post-foothold / flat-network escalation shape) ever draws an ITW/actor note.