al-ice.ai

Six CVEs for an Unknown Number of Bugs: Cisco’s October NX-OS Hardening Release and the Frontier-Model Footnote

AI Vulnerabilities & CVEs by al-ice.ai Editorial

On 7 October 2026, Cisco PSIRT published the advisory batch it had pre-announced a week earlier, covering NX-OS, the Application Policy Infrastructure Controller (APIC), Meraki device software, License On-Prem and Finesse. Four of those advisories are a format Cisco now calls a security hardening release, and they are worth reading closely — not because of any single bug, but because of what the format does to the CVE record.

A hardening release does not describe a vulnerability. It describes a vulnerability class. Cisco states the method plainly in each one: “To assist customers in patching and streamline the disclosure process, Cisco has grouped these issues by their underlying vulnerability class — Common Weakness Enumeration (CWE) — and assigned a single Common Vulnerabilities and Exposures identifier (CVE ID) to each CWE grouping.”

So CVE-2026-76459 is not a bug. It is every out-of-bounds write the NX-OS engineering team found in its internal review, collapsed into one identifier. How many is “every”? The advisory does not say. Neither does the CVE record Cisco filed, which carries the same boilerplate text.

What the scores actually mean

The NX-OS hardening advisory (cisco-sa-hardening-nxosw1-cWzSbtR) lists six CVE IDs and six CWE buckets:

  • CVE-2026-76453 — 8.8 — CWE-707, improper neutralization (command, OS and argument injection)
  • CVE-2026-76455 — 9.8 — CWE-284, improper access control (authorization, authentication, privileges, bypasses)
  • CVE-2026-76456 — 8.6 — CWE-20, improper input validation (including path traversal)
  • CVE-2026-76457 — 8.6 — CWE-125, out-of-bounds read
  • CVE-2026-76458 — 8.6 — CWE-703, improper handling of exceptional conditions
  • CVE-2026-76459 — 9.8 — CWE-787, out-of-bounds write (stack, heap, classic overflow, buffer size miscalculation)

Cisco is candid about the scoring rule, and it is the detail defenders need most: “The CVSS score that is assigned to each CVE ID represents the maximum potential severity of the single most impactful underlying vulnerability within that specific CWE category.”

That is a defensible choice — patch to the worst case — but it breaks the usual inference chain. A 9.8 normally tells you something about one reachable bug. Here, 9.8 tells you that at least one of an unstated number of access-control defects is unauthenticated and network-reachable, and says nothing about the rest. The vector for CVE-2026-76455 is CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H; the vector for CVE-2026-76459 is AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H — 8.8 in the CVE record Cisco filed, against the 9.8 “highest CVSS” the advisory table shows for the same ID. Both numbers are Cisco’s, published the same day, and they are measuring different things: the record scores a representative instance, the table scores the worst instance. Nothing is wrong, but a scanner ingesting one and an administrator reading the other will not agree.

The scope is the part that removes any option to defer. These six affect MDS 9000, Nexus 3000, Nexus 7000, Nexus 9000 in both standalone NX-OS and ACI mode, UCS 6300/6400/6500/6600 Fabric Interconnects and the 9108 100G — regardless of device configuration. There is no feature to disable and no workaround. First fixed releases are 9.4(5a) for MDS 9000; 10.3(10), 10.4(8), 10.5(6) and 10.6(4) for Nexus 3000/9000 standalone; 8.4(14) for Nexus 7000; 16.0(9h), 16.1(6g) and 16.2(3g) for ACI mode; and 4.3(6j) / 6.0(2e) for UCS Manager-managed Fabric Interconnects.

The frontier-model footnote

At the bottom of each hardening advisory, under Source, Cisco writes: “These vulnerabilities were found during internal security testing using existing testing processes as well as frontier AI models.” The same sentence appears in the NX-OS, APIC, Meraki and License On-Prem hardening advisories. The advance-notification document links it to a Cisco PSIRT policy piece titled Strengthening the Foundation: A Predictable, Customer-Focused Response to AI-Accelerated Vulnerability Discovery, and to the company’s transition to a risk-based disclosure model that publishes hardening releases on the first and third Wednesday of each month.

This is the mechanism behind the numbers Google Threat Intelligence described in September when it reported baseline high-risk disclosures roughly doubling over a year. A vendor that can now generate findings faster than it can write advisories has two options: publish a hundred thin CVEs, or publish six thick ones. Cisco chose six, and said so.

The cost lands on anyone downstream of the CVE record. Bucketed IDs do not reconcile cleanly with per-bug workflows: you cannot count them, you cannot tell whether your exposure is one defect or thirty, you cannot track a fix for an individual issue, and a future exploit for one member of the bucket arrives attached to a CVE that was already marked “patched” in your tooling. The compensating detail is that Cisco states these were found internally and are not known to be actively exploited, which is the only reason the format is tolerable — it is an inventory-hygiene exercise, not an incident.

The contrast is in the same batch

What makes the trade-off legible is that Cisco published both styles on the same day. Sitting beside the hardening releases are conventionally-written advisories with real technical content:

  • CVE-2026-76485, -76486, -76501 (cisco-sa-ngoam-rce-LWKQ4BU, CVSS 9.8) — improper input validation in NX-OS VXLAN OAM, known as NGOAM, on Nexus 3000 and 9000 standalone switches. Crafted packets to an IP interface yield arbitrary code execution as root. Cisco spells out the differing preconditions: 76485 needs only NGOAM enabled; 76486 additionally needs SRv6 or NV Overlay with a VXLAN EVPN VNI mapped to an NVE interface and at least one peer VTEP learned; 76501 needs NGOAM plus SRv6. It gives the show feature | include ngoam check, names no feature ngoam as a mitigation, and notes Live Protect shields as a temporary bridge.
  • CVE-2026-76471 (cisco-sa-napi-rce-r2shwu2j, CVSS 9.8, CWE-122) — a crafted HTTP request to the NX-API gives unauthenticated remote root. Cisco notes NX-API is disabled by default on Nexus 3000/9000, and that on UCS 6300 Fabric Interconnects the path runs through the UCS Manager XML API, which is enabled by default and cannot be disabled — but requires valid low-privileged credentials there, so the impact rating drops to High for that platform.
  • CVE-2026-76465 (cisco-sa-moam-rce-uBTzYV7, CVSS 9.8, CWE-590) — a crafted MPLS echo-request to an IP address on an affected device, root code execution again, mitigated by no feature mpls oam.

Every one of those is actionable in a way the bucketed CVEs are not: you can determine exposure, you can mitigate without rebooting, and you can tell a colleague what the bug is. Notably, these three say only “found during internal security testing” — the frontier-model sentence is absent.

The sharpest contrast is in License On-Prem (formerly Smart Software Manager On-Prem), which appears twice. The hardening advisory gives four CWE buckets, topping out at CVE-2026-76482 at CVSS 10.0 for CWE-347, improper verification of cryptographic signature — a perfect score attached to a description with no mechanism in it. The companion advisory cisco-sa-ssm-access-nttb2dhE, covering externally reported bugs in the same product, reads like an advisory should: CVE-2026-20328, an unauthenticated arbitrary account password reset including administrative accounts (9.1); CVE-2026-76454, an unauthenticated API write-anything flaw (9.1); plus an authenticated command injection and a SQL injection. Cisco credits Gabriele Paris of the NATO Cyber Security Centre for three of them and Trung Nguyen of CyStack for the fourth. Outside researchers got named bugs. The internal AI-assisted sweep got buckets.

What to do

  • Treat the hardening CVEs as upgrade triggers, not as risk assessments. There is no configuration to check and no workaround; the only question is which fixed release applies to your platform. Use Cisco’s Software Checker rather than matching version strings by hand, since several platforms have four parallel fixed trains.
  • Patch the feature-gated RCEs first if the feature is on. NGOAM, MPLS OAM and NX-API each gate a 9.8 unauthenticated root path. Run show feature | include ngoam, | include nxapi and the MPLS OAM equivalent across the fleet before scheduling anything — a switch with all three disabled is in a materially different position from one running VXLAN EVPN.
  • Do not let Live Protect shields become the plan. Cisco is explicit that they are temporary mitigations to bridge the gap until an upgrade can be scheduled. They are genuinely useful for a maintenance-window deferral and genuinely not a fix.
  • Check License On-Prem separately and urgently. It carries both a 10.0 bucket and an unauthenticated arbitrary-password-reset flaw, affects all configurations, and is exactly the kind of internal management appliance that gets forgotten in switch-focused patch cycles. Fixed release is 10-202609.
  • Expect your scanner to disagree with the advisory table. Where the CVE record and the hardening table show different base scores for the same ID, the table’s higher figure is the worst-case claim and the record is a representative instance. Patch to the table.
  • Note the Meraki dates before promising a completion date. Two of the Meraki fixed releases are future-dated in Cisco’s own table — 32.2.5 “late Oct 2026” and 33.1.4 “mid-Nov 2026” — so part of that remediation cannot be completed yet regardless of intent.

Verification note: the 7 October 2026 publication date, the advisory list and product groupings, the CWE-bucketing statement, the “maximum potential severity” scoring rule, the six NX-OS CVE/CWE rows and their highest-CVSS values, the affected-product lists, the first-fixed release tables, the NGOAM/NX-API/MPLS OAM technical descriptions and preconditions, the Live Protect language, the Meraki future-dated releases, the License On-Prem bucket table and the researcher credits were all read directly from Cisco’s published advisories on sec.cloudapps.cisco.com. The “frontier AI models” sentence is quoted verbatim from the Source section of the NX-OS, APIC, Meraki and License On-Prem hardening advisories; its absence from the NGOAM, NX-API and MPLS OAM advisories was likewise checked directly. The CVSS vectors and the 8.8-versus-9.8 difference for CVE-2026-76459 were taken from the Cisco-assigned CVE records retrieved from the CVE Program API and cross-checked against NVD, where all of these entries were still in “Received” status at the time of writing with Cisco as the primary scoring source. Cisco states none of these are known to be exploited; we have no independent evidence either way. We tested nothing and exploited nothing.

Sources: