Build1 distinct publisher3 min readPublished
Arista scores CVE-2026-7473 at 5.8 and says it is already being exploited. With every EOS train on the affected list, the thing that decides your exposure is whether a given device terminates a tunnel at all.
The Engineer · Build desk

Compiled by The EngineerSomething wrong?How this is made
The match key is short by one field. On the affected platforms, a packet whose outer destination IP equals the configured decapsulation address gets stripped and forwarded on its inner header, because the switch does not verify the tunnel protocol type [6]. Arista files this as CWE-1023, incomplete comparison with missing factors [4], which is the honest classification: the comparison that decides whether to decapsulate consults the address and stops.
That is why the listener line in the detection output is not a boundary. The example prints "Listening on UDP port 4789" [16], and 4789 is where VXLAN is expected, but a device configured to decapsulate one tunnel type will also accept other tunnel protocols sent to the same IP, with some cases needing extra protocol-specific configuration [11]. Arista adds that the inner header can be IPv4 or IPv6 in every scenario [12], so an IPv4-only service network is not excluded by addressing family.
The scoring is 5.8 under CVSSv3.1 and 6.8 under CVSSv4.0 [2][3], which reads low for something already reported as exploited [7], and the vector is more useful than the number. AV:N/AC:L/PR:N/UI:N is remote, unauthenticated, no user interaction, low complexity. C:N and A:N say the switch neither leaks nor falls over. I:L with S:C puts the integrity damage on a component other than the vulnerable one [19]. In operational terms, nothing breaks on the switch, and a stranger gets one forwarding decision per packet into whatever sits behind the tunnel edge. The reporters are four Comcast engineers, Scott Christiansen, Lukas Peitz, Rich Compton and Jonathan Davis [8], which is an operator's vantage point rather than a lab's.
Now the version list. Trains 4.30.x through 4.36.x are named individually, which is seven, and the list then adds all releases in trains older than 4.30.x and all releases in trains newer than 4.36.x [9]. Seven named plus both open ends leaves no release excluded [18]. A list that closes with everything newer than the current train does save the reader the trouble of checking.
So the discriminator is configuration, and it splits by silicon. Every scenario in the advisory's table applies to the 7020R, 7280R/R2 and 7500R/R2 series [13]. On the 7280R3, 7500R3 and 7800R3, only the IP-in-IPv6 and GUE IPv6 decap group cases apply [14]. One CVE is tracked under two bug numbers, BUG1086442 and BUG1519884 [5], and the advisory does not say which platform group each one covers.
Inventory has to run per device. The command Arista supplies prints "Source interface is <interface> and is active with <IP>" when the box is a VTEP, and that IP is the termination address that makes it potentially impacted [15]. A GRE tunnel endpoint or an ip decap-group puts a decapsulation IP on the box just as effectively [10]. Being out of scope means no EOS device in the estate terminates a tunnel, and that is a claim about running configuration, checked box by box. The version list will not check it for you.
Ranked by verification strength, evidence, and original report placement.
Arista Security Advisory 0137, dated May 5, 2026, tracks the issue as CVE-2026-7473.
CVSSv3.1 Base Score 5.8, vector CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:N/I:L/A:N.
CVSSv4.0 Base Score 6.8, vector CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:L/VA:N/SC:N/SI:L/SA:N.
The advisory classifies the weakness as CWE-1023: Incomplete Comparison with Missing Factors.
The vulnerability is tracked internally by BUG1086442 and BUG1519884.
On affected platforms running Arista EOS where a tunnel decapsulation configuration is present, such as VXLAN, decap-groups, or a GRE tunnel interface, the switch will incorrectly decapsulate and forward other unexpected tunneled packets with a destination IP matching its configured decapsulation IP, because the switch does not verify the tunnel protocol type.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 5, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
build
Arista names four fixed VCO builds for a command injection already in use1 distinct publisher
security
Self-hosted ServiceNow operators inherit three unauthenticated CVSS 10.0 flaws to patch themselves2 distinct publishers
invest
Institutions trimmed the Magnificent Seven and bought chips. That is rotation, not an exit1 distinct publisher
build
Cross-node pod traffic: routing or encapsulation, and why that choice is a debugging decision1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Primary document, single witness
For identifiers, scores and affected trains, the vendor advisory is the authoritative record, and this one is precise down to two internal bug numbers and a per-platform scenario split. The precision stops at the exploitation sentence, which is one clause that gives no traffic, date or reporting party, leaving our coverage with no way to check it.
No exposure counted
This advisory tracks platform families, not device counts. Arista names platform families rather than installed base, there is no fixed release whose uptake could be tracked, and no scanner, catalogue listing or third-party telemetry appears in our coverage. Exposure resolves per site, through a command the reader has to run.
Version list oversells, score undersells
Two distortions pull in opposite directions and do not cancel. An affected list that excludes no EOS release reads as blanket exposure when the actual gate is whether a box terminates a tunnel, which most access switches do not. Against that, 5.8 is a modest number for a defect Arista says is already being used, one that lets outside GRE or IP-in-IP traffic be decapsulated onto internal segments the tunnel type was never meant to reach.
The defect's owner sets the score
Arista is describing a flaw in its own forwarding silicon behaviour, choosing the severity number, deciding how much of the exploitation report to show, and publishing no fixed version. The counterweight is small but genuine: the finders are named engineers at Comcast, a customer with reason to want the behaviour written down plainly rather than softened.
Firm on mechanism, soft on incident
The mechanism, the precondition and the checks are as solid as a vendor advisory gets, and they are the parts an operator acts on. Confidence drops on everything about the attack: who saw it, when, and against which platform. Any fix schedule is unassessable while no release number exists.