

The software supply chain attack that never touched the origin
From 15:01 to 20:30 UTC on 14 September 2026, according to Brevo's own write-up, a malicious Cloudflare Worker rewrote what visitors received while the origin files stayed untouched. That single detail is why so many checks stayed quiet.
By Katie Delaney / 2026-09-20 / 16 min read

How a software supply chain attack got past the origin
A fox does not break into the henhouse when it can borrow the farmer's key. That, in short, is the Brevo incident: a software supply chain attack in which the quarry was never the vendor's servers but the trust that thousands of websites had placed in the vendor's scripts. BleepingComputer reported it on 17 September 2026, and Brevo, the email and marketing platform, published its own account in a status-page write-up.
The root cause of this software supply chain attack, in Brevo's words, was a long-lived Cloudflare API key with full account permissions that had been stored in application source code and was obtained by the attacker. With it, the attacker could create Workers, routes and DNS records on Brevo's zones without triggering an alert. A Worker is a small piece of code that runs at Cloudflare's edge, and Cloudflare's Workers routes documentation says routes let users map a URL pattern to a Worker, so the code runs on matching requests before they reach an application server.
Here is the clock for the software supply chain attack. Brevo's write-up puts the start of impact at 15:01 UTC on 14 September 2026, when the Worker was routed to all of brevo.com, and the end at 20:30 UTC, when the malicious Worker and its routes were removed. That is 5 hours and 29 minutes. At 16:07 UTC, the same write-up says, the Worker was updated to append a loader to three embedded JavaScript files and routed to sibforms.com, which is where customer sites entered the picture. BleepingComputer gives the narrower window of 16:07 to 20:30 UTC, and Sansec logged malicious code from 16:05:18 to 20:12:53 UTC. Three vantage points, three slightly different clocks, one long afternoon.

What did visitors see? Brevo describes a full-screen page dressed as a Cloudflare check that asked the visitor to press Win+R, then Ctrl+V, then Enter, which ran a command placed on the clipboard by the script. That is the ClickFix trick. The page was shown selectively, so most visitors and repeat visits saw nothing. On WordPress sites that embedded a Brevo widget, if the visitor happened to be a logged-in administrator, the script tried to silently install and activate a plugin. Sansec adds that the code did not activate for crawlers, developers or automated scanners, which is the vulpine sort of patience that keeps a quiet raid quiet.
Now the number everyone repeats about this software supply chain attack. Sansec says up to 100,000 sites may have been impacted, and BleepingComputer carries that as Sansec's estimate. Treat it as an estimate of exposure, not a count of confirmed infections. Sansec's write-up does not show the arithmetic, so the fair reading is a very large tail of sites that loaded Brevo components on the day, of which some unknown fraction served the overlay to a real visitor.
Brevo's account of what was untouched is equally specific: app.brevo.com, the Brevo API, email sending and customer account data held in Brevo were not affected. Sansec's note that Brevo disclosed a separate security incident on 10 September is a reason to keep the two stories apart in your own incident notes, and this piece does not treat the 10 September event as part of the 14 September mechanism.
Why the checks you run missed it
Why did the usual tripwires stay still during a software supply chain attack this quiet? Brevo's answer is blunt: because the Worker rewrote responses at the edge and removed security headers such as Content-Security-Policy, its origin servers and files remained unmodified and standard integrity checks did not detect the change. A file-hash monitor that watches the origin is watching the wrong den. The attacker sat in front of the door, not inside the house.
Two browser controls are worth reading against that story, because both get sold as cure-alls. The W3C Subresource Integrity specification defines a mechanism by which a browser checks a fetched file against a cryptographic hash the site author expects, and it says compromise of a third-party service should not automatically mean compromise of every site that includes its scripts. The W3C Content Security Policy specification describes a policy, preferably delivered as an HTTP response header, that mitigates content injection. Both are real and both are useful. Neither is a force field.
Take integrity pinning first. A pinned hash would have refused a changed file, but only on sites that had pinned it, and a vendor loader that ships updates on the vendor's schedule is precisely the file most sites cannot pin without breaking on release day. That is folkfox analysis, not a Brevo statement, and it is why pinning protects the static libraries you host and rarely the widgets you rent.
Then the policy header. Here the evidence is more interesting than the slogan. The second-stage script came from a different host, a domain Brevo owns called sendibt1.com, and Sansec reports that its CSP monitor recorded 2,549 violation reports across 12 sites in and after the window. Read that as a tripwire working for a handful of sites that had a policy naming hosts and a reporting endpoint. It says nothing about the other tens of thousands, and it would say nothing at all if the attacker had served the code from a host you already allow. A policy lists names, not intentions.
| Control | What the source says it does | Against this incident | Usual owner |
|---|---|---|---|
| Integrity hash pinning | Browser checks a fetched file against a hash the author expects (W3C) | Refuses a changed file where pinned; loaders rarely pinnable | Web engineering |
| Content Security Policy | Header-delivered policy that mitigates content injection (W3C) | Flags new hosts; trusted hosts serving altered code pass | Security or engineering |
| Scoped, expiring tokens | Cloudflare tokens can be limited by permission, zone, client IP and TTL | Full account permissions, long-lived, by Brevo's account | Vendor platform team |
| Secrets scanning | Finds credentials sitting in code before an attacker does | Key sat in source code; now short-lived tokens, says Brevo | Vendor engineering |
- Integrity hash pinningBrowser checks a fetched file against a hash the author expects (W3C)Refuses a changed file where pinned; loaders rarely pinnable Web engineering
- Content Security PolicyHeader-delivered policy that mitigates content injection (W3C)Flags new hosts; trusted hosts serving altered code pass Security or engineering
- Scoped, expiring tokensCloudflare tokens can be limited by permission, zone, client IP and TTLFull account permissions, long-lived, by Brevo's account Vendor platform team
- Secrets scanningFinds credentials sitting in code before an attacker doesKey sat in source code; now short-lived tokens, says Brevo Vendor engineering
The row that matters most is the third. Cloudflare's API token documentation shows how a token can be restricted, for example that read access to DNS on one zone lets the token read DNS records only for that zone, with optional client IP filtering and a time to live. A key built that way is a small key. Brevo's write-up says the replacement is narrowly scoped, short-lived tokens, held in a vault with automatic rotation, which is the same lesson in the past tense.

Which supply chain attack example should you study?
| Item | Value |
|---|---|
| Third party involved | 30 to 48 |
| Exploited flaw as entry | 20 to 31 |
Is this software supply chain attack an outlier? The measured picture says no, with caveats. Verizon said on 19 May 2026 that breaches involving a third party now account for 48% of all breaches, and that vulnerability exploitation, at 31%, has for the first time in 19 years overtaken stolen credentials as the top way in. Verizon's 2025 report , published on 23 April 2025, put third-party involvement at 30%, which Verizon described as a doubling, and exploitation at 20%. Third-party involvement is a wide net that covers far more than script injection, so it is the climate, not the forecast for one site.
The formal vocabulary has been around for a while. NIST SP 800-161 Revision 1, published in May 2022, gives guidance on identifying, assessing and mitigating cybersecurity risks throughout the supply chain, and CISA and NIST guidance on defending against software supply chain attacks sits alongside it. What none of it can do is tell you which of the forty scripts in the thicket on your checkout page is the one that goes bad on a Monday.
So, which supply chain attack example should a marketer keep in the drawer? Three are properly verified. The first is Brevo, on 14 September 2026. The second is the polyfill.io case of June 2024, in which, according to Sansec, a Chinese company that acquired the polyfill domain and its GitHub account in February 2024 then had the domain injecting malware on mobile devices via any site that embedded it, on more than 100 thousand sites. The third is the npm worm that CISA's September 2025 npm alert described on 23 September 2025, which compromised over 500 packages and hunted for developer credentials and cloud API keys.
Notice the rhyme, because every software supply chain attack of this shape has one. Polyfill was a trusted domain changing hands. npm was a developer credential turned into a spreading tool. Brevo was a single key with a very long reach. In each, the victim organisations did nothing unusual, which is what makes them so unfair and so instructive. We took the banking-side view of vendor dependence in our piece on third party risk management guidance, and the same logic applies to the scripts your marketing team installs on a Friday afternoon.
Stolen Cloudflare API key (not a vulnerability in Brevo's platform itself).
That one sentence, from a security news forum three days after the write-up, is the whole moral of this software supply chain attack in eleven words. Nobody found a clever hole. A door key was lying where keys should not lie, and the door was the one thousands of sites had been told to trust.
Martech security is a marketing problem
Here is folkfox's opinion, and it will irritate a few people on both sides of the org chart. Martech security is a marketing problem wearing a security badge. The scripts behind this software supply chain attack were tracker loaders, chat widgets and form embeds, the sort of thing a marketer adds through a tag manager in ten minutes and a security team learns about after an incident. Ownership follows the budget, the trail of invoices leads to marketing, and the undergrowth of tag-manager children grows there unobserved.
That has three practical consequences. First, the inventory of third-party scripts is a marketing deliverable as much as a security one, so put it in the same spreadsheet as the campaign calendar. Second, every new martech vendor needs a two-line security answer before contract, not after. Third, the person who can switch a widget off in a hurry must be named in advance, because an incident at 19:33 UTC will not wait for a steering group.
Are your cloud and CDN credentials scoped to the least permission and zone needed, or does any key hold full account rights?
How long does a credential live, and is rotation automatic? Brevo's remediation names a vault with automatic rotation.
Do you scan code and configuration for credentials before they ship, and who reads the alerts?
Does every change to Workers, routes, DNS or account access raise an alert a human reviews? Brevo now says it does.
If you were compromised tomorrow, how fast could you tell customers the exact window and the exact check to run?
Those questions are cheap, and vendors who cannot answer them in a week are giving you an answer. If your own remit is agent credentials and automation, the same discipline appears in our piece on agentic AI security and the runaway agent, and the identity side of it is in passkey phishing and cloud identity protection. A key is a key, whichever species of software holds it.
For security vendors and MSSPs there is a second, colder point. Your own marketing stack is in scope of your own credibility the day the next software supply chain attack lands. A cybersecurity firm that loads a dozen unmanaged scripts on its homepage is telling prospects something without meaning to. That is a theme we return to in cybersecurity marketing, where the trust signal is often a page that does not misbehave.

Supply chain attack prevention for the stack marketers own
| Item | Value |
|---|---|
| Set-up to impact | +38 (running total 38) |
| Unseen exposure | +272 (running total 310) |
| Alarm to removal | +57 (running total 367) |
| Set-up to end | 367 |
Look at the middle bar. The attacker's set-up took 38 minutes from the first hostname at 14:23 UTC to the start of impact at 15:01. Then 272 minutes went by before anyone opened an incident. Once the incident was open, removal took 57 minutes, and an independent check confirmed clean pages 12 minutes after that. The hard part was not the fix. It was the noticing, and the notice-to-fix ratio is the number to defend in a boardroom.
That shapes supply chain attack prevention for a marketing team facing the next software supply chain attack, which is mostly about shrinking the blast radius and speeding up the noticing. Begin with a patient prowl through the inventory. Load fewer third-party scripts, and delete the ones nobody can explain. Where a vendor asset is static, pin it with an integrity hash. Where it is not, run a Content-Security-Policy in report-only mode first, with a reporting endpoint, so an unexpected host shows up in a log, a scent on the trail, before it shows up in a headline. Ask for a per-vendor kill switch, and rehearse it once.
A patch reminder tells readers you exist. A were-you-exposed answer tells them you can count.
That pullquote is the commercial point. In the hours after a software supply chain attack, buyers do not want a webinar. They want the window, the indicators and the check. Brevo's write-up gave times to the minute. Sansec gave a log search for a POST to /wp-admin/update.php?action=upload-plugin on the day, and a look at plugin install dates. A security vendor or MSSP that publishes that sort of specific, checkable answer within a day earns the attention that a patch reminder never will, and it earns it from exactly the accountable buyers who sign the pipeline.
If that sounds like content strategy, it is. Our content marketing services and SEO and GEO work start from the same test: the page that answers the sharp, time-boxed question wins. We tested the pattern against regulated timelines in the 24-hour credibility test of the CRA reporting clock, and against patch cycles in Microsoft security updates patch triage and SAP patch day response.
One caution on fairness. Brevo's write-up is Brevo's account, and Sansec's reading of a Cloudflare account compromise is Sansec's inference from DNS and header evidence, not a confirmed forensic finding. The write-up says the key was first misused in late August 2026 and that Brevo found no injection into customer-facing pages before 14 September. Take the numbers as reported and hold the causal story loosely until an independent review lands.
If you run a martech stack, or sell security to people who do, talk to folkfox about turning a dull scripts inventory into a trust signal buyers can see.
Frequently asked questions
What are some recent supply chain attacks?
Three verified examples of a software supply chain attack: Brevo on 14 September 2026, where a leaked Cloudflare key let an attacker rewrite scripts at the edge; the polyfill.io domain injecting malware in June 2024, per Sansec; and the npm worm CISA described on 23 September 2025, which compromised over 500 packages.
What is a software supply chain attack?
It is an attack that compromises code, a service or a credential that a target already trusts, rather than the target itself. The victim loads the attacker's code through a normal, approved route, which is why ordinary defences often stay quiet.
Was Brevo customer data stolen in the incident?
Brevo says app.brevo.com, its API, email sending and customer account data were not affected. The exposure it describes is on the visitor side and, for WordPress sites, in administrator sessions, so customers should still check plugin installs from 14 September 2026.
How many sites were affected by the Brevo incident?
Sansec estimates that up to 100,000 sites may have been impacted, as BleepingComputer reported. That is an estimate of sites embedding affected Brevo components, not a confirmed count of infections, and Brevo has not published a customer figure.
Would Subresource Integrity or a Content Security Policy have stopped it?
Partly, in specific cases. A pinned integrity hash would refuse a changed file, but few sites pin self-updating vendor loaders. A host allow-list could flag the attacker's separate loader host, and Sansec logged 2,549 policy violation reports across 12 sites.
What does supply chain attack prevention look like for a marketing team?
Keep an inventory of third-party scripts, remove what nobody can explain, pin static assets, run a reporting Content Security Policy to spot unexpected hosts, name a person who can switch a widget off fast, and put key scoping and rotation questions into vendor onboarding.
Cybersecurity marketing now has a 24-hour credibility test
How to explain EU reporting duties without turning compliance into a badge.
Read the pieceThird-Party riskThird party risk management: the proposed reset banks should not sell as settled
Proposed, nonbinding guidance and four honest moves for bank and fintech marketers.
Read the pieceAgentic AIAgentic AI security and the fox that watched the meter climb
Governing spend and access before an agent acts, for MSSPs and vendors.
Read the pieceIdentityPasskey phishing and the polite path to cloud compromise
Five protection moves for teams defending cloud access.
Read the pieceBe the vendor that answers first
folkfox builds cybersecurity marketing that publishes the specific, checkable answer in the hours after an incident, and audits the martech stack behind it.
Want folkfox in your Google results and AI answers? Set folkfox as a preferred source.

