top of page

Misplaced Trust: This Week's Exploited Bugs Were a Median 640 Days Old, but the Victims Mostly Got In Through a Vendor, an Admin Login or a Box Nobody Owned

Writer: Patrick Duggan
Patrick Duggan
12 minutes ago
7 min read

If you feel buried by the pace of patches, here is the measured version of this week. Of 187 vulnerabilities with first exploitation evidence dated November 13 through 20, the median one had been public for 640 days. More than half were over a year old. At most 6% were zero-days. The flood of CVEs is real, but attackers spent this week feeding on old bugs.


Then we looked at who actually got breached this week, and the picture changed. Of 14 compromises with a named victim disclosed November 14 through 20, half never said how the attacker got in. Of the seven that did, four came in through a vendor, one through a stolen admin credential, one through a fix that sat unapplied, and one through two chained zero-days. Only one of the seven was a patching failure.


The armor is not the problem. The soft spots are between the plates, and every one of them is misplaced trust: a vendor trusted to write into your pages or your network, an admin login trusted because someone tapped Approve, a box trusted to keep working after its maker stopped fixing it, an advisory trusted to list the right versions. Trust is not a control. Each item below turns one of those trusts into something you have checked.



What to do this week


1. List every vendor that can write to you, and treat that list as your attack surface. That means every script vendor your web pages load (testing, analytics, tag managers, chat widgets), your managed IT provider, and your remote management tools. Four of this week's seven known entry points were a vendor: Optimizely's snippets served a skimmer and a crypto clipper to 178 organizations, a breach at Fireworks AI reached Uber's driver data, a marketing vendor exposed Visa, Mastercard and JPMorgan customer records, and an IT provider exposed about 40 Irish credit unions. None of those victims could have patched their way out. What they could have known in advance is which vendor could touch what.


2. Put phishing-resistant or number-matching MFA on every super admin, today. COLD RELIC took a second US organization this week with "the stolen credentials of a Super Admin account" and an MFA push-fatigue attack, per Google. A push prompt that can be approved by a tired person is not a control for an account that can change everything.


3. Retire what is past end of support, on a date. SonicWall said this week that Gen 6 firewalls will get no more security patches, even for critical flaws, and the last Gen 6 models left support on October 1. Of the 187 exploited bugs this week, 39 were five or more years old, many in software whose maintainer is gone. For those, the patch is removal.


4. Patch from the exploited list, not the CVE firehose. CISA KEV added seven entries this week, with a median of 14 days between disclosure and listing. VulnCheck's own canaries caught 38 more being exploited, with a median of 81 days. That is a short list, and every item on it already had a fix or a retirement path. If your queue is ordered by CVSS instead of by "is this being exploited and do we run it", you are working the wrong list.


5. Take management planes off the internet. This week's one true zero-day chain, ARMORY CASTLE against vCenter, could not have been patched in advance. It ran over vCenter's directory service on TCP 2012, after entering through a BeyondTrust zero-day. The Cisco Expressway break-in ran over cluster ports 4371-4372. Neither port needs to be reachable by strangers. Patching speed loses to a zero-day; a firewall rule between strangers and an admin port does not care whether the bug has a name yet.



The numbers


Population 1: exploited bugs. Every CVE whose first exploitation evidence is dated November 13 through 20, 2026: VulnCheck KEV entries (VulnCheck dates each entry to its earliest exploitation evidence) plus CISA KEV additions in the same window. 187 CVEs. We measured the time from public disclosure to that first evidence.


  • At or before disclosure (the ceiling on zero-days): 12, or 6%

  • 1 to 30 days: 37, or 20%

  • 31 days to 1 year: 35, or 19%

  • More than 1 year: 103, or 55%

  • More than 5 years: 39, or 21%

  • Median: 640 days

Two subsets are worth reading on their own. The seven CISA KEV additions are the fast lane: median 14 days, none at disclosure, all with a fix available. The 38 that VulnCheck's own canaries saw exploited are observed attacks rather than list entries: median 81 days, 63% older than a month.


Population 2: victims. Every compromise with a named victim disclosed November 14 through 20, 2026, that we or the outlets we read covered. 14 incidents. Leak-site claims without a confirmed breach (The Gentlemen's NVIDIA claim) are left out, as are researcher demonstrations (Kokoro, the Docker Hub decoy keys).


  • Entry not disclosed: 7 (RIKEN Fugaku, N-able Mail Assure and SpamExperts, SBC Medical, TfL Oyster photocards, Fandango checkout, the EggStreme v2 government target, Microsoft's agent-framework PyPI package, where Wiz suspects but has not confirmed a stolen release credential)

  • Through a vendor or third party: 4 (Optimizely's 178 customers, Uber via Fireworks AI, Visa, Mastercard and JPMorgan via a marketing vendor, Irish credit unions via Wellington IT)

  • Stolen credential: 1 (COLD RELIC's second US victim)

  • Fix available, not applied: 1 (a Russian organization on Cisco Expressway X15.5.3, about two weeks or more after Cisco's October 21 fix, per F6)

  • Zero-day: 1 (ARMORY CASTLE, vCenter CVE-2026-123890 since early September, entered through BeyondTrust CVE-2026-102764)


Where this argument is weak


We would rather show you the counterexamples than have you find them.


  • Half the victims never said how. Seven known entry points out of 14 is a small sample. The undisclosed half could be all zero-days. It probably is not, but we cannot show that.

  • Some victims did their jobs and lost anyway. N-central servers mining Monero on November 12 (just outside this window): 19 of the 21 were already running N-able's HF2 fix, and new public exploits claim to beat it. Those operators patched.

  • The fix-to-attack window is shrinking for internet-facing software. UnrealIRCd was fixed November 7 and attacked November 8. Puppet Enterprise went from fix to CISA KEV in 14 days. Google described attackers using AI agents to build exploits from patches. For an internet-facing box, "a fix existed" now often means days, not months.

  • Sometimes the vendor tells you wrong. Apache Roller's first advisory for CVE-2026-129448 named one version. Roller 4.0.0 through 6.1.4 were exposed all along, and it is being exploited. Admins who read the advisory and moved on did what they were told.

  • "Old" does not always mean a fix was sitting there. Many of the oldest exploited bugs are in abandoned products: dead WordPress plugins, unmaintained CMS projects. The fix there is retirement, which is a lifecycle job, not a patching one.

So the fair reading is not "defenders are slow." The pace of new bugs was never what took people down this week. What took them down was the seams: a vendor, an identity, a box nobody owned. Most teams have never been handed a list of those seams, and that is a method problem, not a people problem.



Am I affected? A self-check for the seams


  • Vendors: open your homepage and checkout page in a browser, view the page source, and list every external script host. For each one, name the vendor and who in your organization owns that relationship. A host nobody can name is your first finding. To check whether any host on your homepage is on our list, run Am I affected?. It reads your homepage HTML only. It catches a bad host written into your own page, and it cannot see code injected inside a vendor's own script file, which is how the Optimizely wave worked. A clean result means "nothing on our list", not "safe".

  • Admin identities: in Google Workspace, Microsoft Entra or Okta, list every super admin or global admin and the MFA method each uses. Any account on push approval without number matching, or on SMS, is a finding.

  • End of life: for each internet-facing appliance, write down the model, the firmware version and the vendor's end-of-support date. Anything past that date with a public IP is a finding.

  • Exploited and in your estate: take this week's CISA KEV additions and VulnCheck's canary-confirmed list, and search your asset inventory for each vendor and product. If you cannot run that search, the missing inventory is the finding.

  • Admin ports: from outside your network, check whether vCenter (TCP 443 and 2012), Expressway cluster ports (4371-4372), or any firewall or VPN management page answers. Each one that answers is a finding.


Block these


Where a primary source publishes indicators, we put them in the feed the same day. Today that was 11 ARMORY CASTLE obfuscation-network IPs from Google's report and 2 Cisco Expressway attacker IPs from F6, each verified after ingest. Get the feed (free registration) and read the docs to wire it into your firewall or SIEM. A blocklist is armor on the plates. It does not cover the seams above, and nothing we sell does. Those take a list only you can make.



How we know


  • The exploited-bug population comes from VulnCheck KEV (entries with date_added November 13 through 20, 2026) and the CISA KEV catalog (version 2026.11.18, downloaded November 20). Disclosure dates are NVD publish dates, read through VulnCheck's NVD index.

  • 66 of the 187 CVEs are old bugs that received a CVE number only recently, so their NVD publish date is new. For those, we dated disclosure no later than December 31 of the year in the CVE number. Their ages are lower bounds; the real ages are likely higher.

  • Disclosure date stands in for fix date. For most bugs the two are close. For abandoned software there is no fix date at all.

  • The victim census comes from our own posts this week, each built on the primary source, which is linked from that post. A class was assigned only where the victim or a named researcher disclosed the entry. Everything else is counted as not disclosed, not guessed.

  • VulnCheck KEV data is VulnCheck's. A VulnCheck date is its earliest exploitation evidence, not the date it listed the bug.


Sources and our posts on each incident






Every indicator in this post is in the feed. Free.

1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.


Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page