3
Connect Secure CVEs in the chain to root
8.6
CVSS for the entry bug, CVE-2026-104935
16 days
between the entry fix and the root fix
0
web shells or backdoors left on the appliance
On an edge VPN, patched does not mean clean
The Ivanti Connect Secure appliance that Volexity investigated in early October was fully patched when the attack started, and it stayed that way while an attacker ran commands on it as root. No web shell was left behind. Ivanti's own integrity checker found nothing, because the exploit changed code in memory and put the configuration back the way it found it.
That is the argument of this post. On an internet facing VPN, the patch level answers one question: what can happen to this box from today on. It says nothing about what already happened. And on Connect Secure the thing worth stealing is not the appliance at all. It is the directory credentials the appliance holds so it can check your users' passwords.
CISA closed the loop on 30 Oct 2026 when it added CVE-2026-111357 and CVE-2026-118527 to its exploited catalog, the last two links of a three bug chain whose first link, CVE-2026-104935, was listed and patched on 13 Oct. For about two weeks the advice was to patch the entry bug. The privilege escalation to root was not fixed until 29 Oct, 16 days later. An appliance that took the first patch promptly was still running a build that CISA now lists as affected.
We have made versions of this point before, with a Cisco firewall bug that ends sessions and with stolen VPN session cookies on Palo Alto gateways. This case is cleaner, because the investigator published what the attacker did after getting in, and almost all of it happened somewhere other than the VPN.
What is the Ivanti Connect Secure CVE-2026-111357 chain?
Three CVEs make up the chain. CISA's catalog describes each one, and its own text says each can be chained with the other two.
CVE-2026-104935 is the entry point. CISA describes it as an authentication bypass by spoofing that lets "a remote unauthenticated attacker" execute arbitrary commands. Rapid7 gives it a CVSS 3.0 score of 8.6, a high on our scale, and The Hacker News reports the same 8.6 as Ivanti's rating. Ivanti shipped fixes on 13 Oct, and CISA listed it the same day with a three day deadline.
CVE-2026-111357 is the step to root. CISA calls it a missing authentication for critical function flaw that lets a local authenticated attacker escalate privileges. Volexity locates it in how a privileged broker process on the appliance trusts cached configuration data, which lets an attacker run commands as root. Ivanti published the patch on 29 Oct.
CVE-2026-118527 is a predictable identifier flaw. CISA says it lets a remote authenticated attacker upload files with attacker controlled content to a known location on disk. CISA added both of the newer CVEs on 30 Oct. The deadline for CVE-2026-111357 is 2 Nov, with forensic triage required under BOD 26-04. The deadline for CVE-2026-118527 is 13 Nov, with no triage requirement.
Read the two "authenticated" descriptions with the first bug in mind. An attacker who has already bypassed authentication and is running code on the box meets the precondition. That is what chaining means here.
-
Entry //
CVE-2026-104935, unauthenticated, CVSS 8.6 (3.0), fixed 13 Oct -
Root //
CVE-2026-111357, privilege escalation, fixed 29 Oct -
Upload //
CVE-2026-118527, predictable file locations -
Affected, per CISA //
25.1.1.2 and prior, 25.1.2.2 and prior, 22.8R2.5 and prior -
Exploited //
Yes, all three in CISA KEV
The version list is where operators get caught. Rapid7 lists 22.8R2.5 and 25.1.1.2 among the builds Ivanti said resolved the entry bug in mid October. CISA's alert, last revised on 30 Oct, lists those same builds inside the affected range for the chain. A team that upgraded to 22.8R2.5 on 13 Oct did the right thing that day and is still exposed today. Check Ivanti's advisory for the build on your code line that carries the 29 Oct fix, and do not trust a ticket that closed in mid October.
Ivanti said at disclosure that it was "aware of limited exploitation in the wild." Volexity described one compromised appliance at a customer it monitors. Neither has said how many organizations were hit, and we will not guess.
How three bugs became root without a login
Volexity's write up describes the chain as two flaws that give code execution as the low privileged web server user, followed by a privilege escalation to root. In plain terms, the attacker went from an anonymous request to the most powerful account on the appliance, and never typed a password.
Figure 01
From the login page to the directory
- 01
Bypass //
An unauthenticated request runs code as the web user.
- 02
Root //
A trusted broker runs the attacker's commands as root.
- 03
Harvest //
Stored domain credentials leave the box.
- 04
Directory //
LDAP reads for LAPS and BitLocker secrets.
- 05
Erase //
Debug and audit logs lose the evidence.
Result //
A patched VPN became a key to the domain.
The interesting part starts at step three. Volexity reports that UTA0576 extracted credentials stored on the appliance, "such as domain credentials used with LDAP." Connect Secure, like most VPNs, binds to Active Directory with a service account so it can authenticate users and read group membership. That account usually has read access across the directory, and its password sits on the appliance in a form the appliance itself can use.
With that account, the attacker ran LDAP searches for Local Administrator Password Solution passwords and BitLocker recovery keys. Volexity also saw password spraying and a tool capable of Kerberoasting. Those are ordinary internal network moves. What is unusual is where they came from: the VPN's own address, using the VPN's own account.
Volexity assesses with low confidence that UTA0576 has criminal rather than espionage motives, and with medium confidence that the actor uses large language models. We repeat both assessments at the confidence Volexity gave them. Neither changes what a defender should do this week.
Why the integrity checker found nothing
Ivanti ships an Integrity Checker Tool so customers can ask the appliance whether its files have been tampered with. Volexity's explanation of why it failed here is one sentence long: the chain "patches library code in memory only and restores configuration store data," so the tool "does not detect any modifications." CISA's alert repeats the warning, saying the checker may not detect exploitation because the commands execute in memory.
Put that next to the rest of what Volexity found. There was no backdoor and no web shell on the device. Volexity says UTA0576 deleted evidence of exploitation from Ivanti debug logs and Linux audit logs. The actor did not install persistence because it did not need any: it could repeat the exploit whenever it wanted access again.
That pattern defeats the usual post patch routine. The routine goes like this: upgrade, run the integrity check, look for files that should not be there, close the ticket. Every one of those steps assumes the attacker left something on the box. This one left none.
There are traces, but they are thin. Volexity published the network pattern it saw, a high volume of requests to a Connect Secure login page carrying a request header whose path did not match the page, along with files written under temporary directories and an address the actor used. Those are worth searching for in proxy, load balancer and network monitoring data you hold outside the appliance. Data held on the appliance is the data the actor was editing.
It is also worth being honest about the 16 days between the two fixes. During that window a fully updated appliance was still open to the step that turned web server access into root, and the public guidance was to patch the entry bug and watch for the rest. Teams that patched on 13 Oct and stopped looking had no way to know whether the entry bug had been used against them before they patched. That question does not expire when the second patch lands. It has to be answered with evidence from outside the box.
The stronger signal is downstream. Whatever the attacker does with the credentials happens on systems you control and log: domain controllers, file servers, identity providers. That is where we would put the effort.
Where to hunt: the directory, not the appliance
Start with the VPN's LDAP bind account. In normal operation it does a narrow, repetitive job: bind, look up a user, read groups. It has no business reading LAPS passwords or BitLocker recovery keys. If it did, someone was using it who was not the VPN.
The attribute names are documented by Microsoft. Windows LAPS stores the current password in msLAPS-Password or msLAPS-EncryptedPassword, legacy LAPS in ms-Mcs-AdmPwd, and BitLocker stores recovery passwords in msFVE-RecoveryPassword on msFVE-RecoveryInformation objects. A query that names any of them, from the VPN account or from the VPN's internal address, is worth a person's attention.
Figure 02
VPN bind account reading LAPS or BitLocker secrets
KQL
// Replace with your Connect Secure LDAP bind account names
let VpnBindAccounts = dynamic(["svc-vpn-ldap"]);
let SecretAttributes = dynamic(["ms-Mcs-AdmPwd", "msLAPS-Password",
"msLAPS-EncryptedPassword", "msFVE-RecoveryPassword",
"msFVE-RecoveryInformation"]);
IdentityQueryEvents
| where Timestamp > ago(30d)
| where Query has_any (SecretAttributes)
| extend FromVpnAccount = AccountName in~ (VpnBindAccounts)
| summarize Queries = count(), Targets = make_set(QueryTarget, 20),
Sources = make_set(IPAddress, 10)
by AccountName, FromVpnAccount, bin(Timestamp, 1h)
| order by FromVpnAccount desc, Queries desc Run it for at least the window since 3 Oct, the date Volexity gives for the start of exploitation. Sort the VPN account to the top, then look at everything else: a help desk account reading BitLocker keys is normal, a print server is not.
Two more hunts belong next to it. First, look for password spraying sourced from the VPN appliance's internal address: many accounts, few attempts each, in a short window. Our password spraying page covers what that looks like from the attacker's side. Second, look for service ticket requests for accounts with service principal names coming from the same address, which is the trace Kerberoasting leaves. Neither is exotic, and both are things a VPN should never do.
If any of these come back positive, treat the appliance as compromised whatever its version says, and move straight to recovery.
How we test an edge VPN on an engagement
We start by finding every VPN login page the client exposes, not just the one in the scope letter. Our external recon method maps host names from DNS, certificate transparency and the addresses in scope, and our open source atarus-recon tool automates part of it. Forgotten portals on old builds are a common finding.
On a sanctioned assumed breach test we also start from the other end, with the question this incident raises: if the VPN were ours, what could we reach with what it stores? That means reading which account the appliance binds to the directory with, what that account can read, and whether anyone would notice it reading things it never normally touches. Our internal recon page walks the directory questions we ask in the first hour.
The finding we write up is usually not the VPN bug. It is the gap between the VPN account's real permissions and its job. A bind account that can read every LAPS password is a domain admin path with an extra step. Fixing that takes an afternoon of delegation work in Active Directory and holds up through every future appliance bug, including the ones nobody has found yet.
The same review covers what else the appliance holds. Certificates for client authentication, API keys for management tools, and credentials for license servers all sit on Connect Secure, which is why CISA's recovery list names each of them. We list every secret on the box, who could use it, and what it opens. That list is the scope of the rotation work after an incident, and it is far better written before one.
We also test whether the client would see the hunts above fire. A detection that has never been tested against real activity is a hope, not a control.
What is still coming
CISA's alert notes that Connect Secure 22.8 reaches end of life on 31 Jan 2027. Anyone on that line has three months to plan a move that Ivanti's own guide treats as a migration, not an upgrade. Teams that delay it will be applying emergency fixes to an appliance with a hard stop date, which is the worst time to learn a new code line.
The third CVE deserves attention too. Volexity's write up describes the file write behaviour in the chain but does not tie it to CVE-2026-118527, and none of the pages we read give a separate fixed build for it. Until Ivanti's advisory says plainly which build fixes it on your line, treat the chain as one unit and take the newest build available.
And the shape will repeat. Edge appliances are attractive because they sit on the internet, hold directory credentials, and run closed systems that customers cannot inspect. When we wrote about a week of edge device exploitation in June, the lesson was the same: the device is a doorway, and the room behind it is your directory. The patch closes the doorway. Only recovery tells you whether someone is already in the room.
What to do today
-
Upgrade to the build that carries the 29 Oct fix. Anything at or below 25.1.1.2, 25.1.2.2 or 22.8R2.5 is in CISA's affected list, including builds that fixed the entry bug on 13 Oct.
-
Follow CISA's recovery steps, not just the patch. Back up the configuration, factory reset and upgrade, rebuild cloud and virtual instances, then restore. The integrity checker cannot clear this chain.
-
Rotate everything the appliance stores. CISA says reset domain, administrator and local account passwords, revoke and reissue certificates, and reset API keys and license server credentials. Start with the LDAP bind account.
-
Hunt the directory since 3 Oct. Run the KQL above for LAPS and BitLocker reads, then look for spraying and service ticket requests from the VPN's internal address.
-
Cut the bind account down to its job. It needs to look up users and groups. It does not need LAPS passwords or BitLocker keys.
-
Plan the move off 22.8 now. It reaches end of life on 31 Jan 2027.
SOURCES
- CISA // adds CVE-2026-111357 and CVE-2026-118527 to KEV on 30 Oct 2026
- CISA // affected versions, the integrity checker caveat and the recovery steps
- Volexity // the chain, UTA0576, the fully patched victim and what the actor took
- Rapid7 // the 8.6 score and the builds Ivanti listed as fixing CVE-2026-104935
- Ivanti // confirms limited exploitation in the wild at disclosure