Skip to main content
All Articles

Security Weekly 2026-W47: The Card Fields You Built Yourself Were the Ones That Got Hit

Category Security Date 2026-11-23

Author
恩梯科技
26 views 19 min read

This is the fifth issue of the NT IT security weekly, covering week 47 of 2026 (from 16 November). Data was read up to the evening of 22 November. It was a full week: a supply-chain injection that reached hundreds of well-known sites, PHP's CRITICAL batch, Laravel and WordPress each emptying a batch of fixes on the same day, plus three flaws in AI tooling itself.

How to read this weekly

In every section we separate three things: what the vendor itself said (verbatim, not paraphrased), what that sentence actually means (our reading, and we mark it as ours), and whether you need to do anything.

One rule has not changed: when we looked and found nothing, we say so. That half usually helps you judge more than the findings do — because you can tell whether we actually looked, or simply left it out.

1. Optimizely: malicious code in the sites of 178 organisations, and whether you were hit depends on how your card fields were built

On 16 November, attackers abused Web Experimentation at the A/B testing platform Optimizely to insert malicious JavaScript into customers' websites.

What Optimizely's own status page said (verbatim):

Earlier today, Optimizely identified a security incident in which an unauthorized party exploited a vulnerability in Experimentation and used it to add malicious JavaScript to some customer projects. We've fixed the vulnerability and removed the injected code from affected projects.

When we looked on 18 November, the company had not yet said how many customers were affected, which ones, how the attacker obtained cross-customer write access, or whether any data left the building. The status was Monitoring.

But when we looked again before publication (the evening of 22 November), the detail had been filled in (verbatim):

On 16 November 2026, an unauthorized third party exploited a flaw in the Optimizely Web Experimentation account signup flow. They used it to impersonate users on targeted customer accounts and insert malicious JavaScript into those customers’ projects. On the day the activity began, we detected it, blocked delivery of the malicious code and closed the vulnerability within three hours of detection. We also removed the injected code from all affected projects. No malicious code remains in our systems. This was not a credentials breach. The incident was limited to Web Experimentation. Less than 5% of our customers were impacted. No other Optimizely product was affected.

So the three blanks now have answers: the entry point was a flaw in the signup flow — the attacker used it to impersonate users, rather than stealing anyone's password (verbatim: This was not a credentials breach); three hours passed between detection and closing the vulnerability; and fewer than 5% of customers were affected. Optimizely also says it has contacted every Web Experimentation customer directly.

The part of this worth remembering is not those numbers, it is the gap in time: the incident was on 16 November, on 18 November the company had said only "we fixed it", and the entry point and scope arrived on 22 November. If you were deciding on 17 November whether to switch that snippet off, none of this was in your hands. (That sentence is our judgement, not the company's.)

The number comes from research, not from the vendor: security researcher Scott Helme reported at Report URI that on 16 November visitors to 178 organisations were served malicious JavaScript, naming customers including Zoom, BBC, CNN, Salesforce, GitLab, IKEA and HSBC. He also wrote down how the figure was produced — they collected 1,611 distinct Optimizely snippets and compared them, of which 188 (belonging to 178 organisations) were infected — and immediately added that the real figure may be higher, because they could only check the ones they managed to find in time. So do not treat 178 as a complete total; that is the researcher's own caveat.

The mechanism is worth understanding, because it explains why "I didn't install anything odd" does not save you: every customer loads a snippet from Optimizely's CDN, and that snippet contains a field the customer fills in themselves. The attacker wrote code into that field, without altering a single character of the rest of the snippet, so the file was still served from Optimizely's domain with Optimizely's certificate. The report states verbatim: this was not a CDN cache-poisoning attack. Optimizely's own website and login page were hit too.

The injected code was doing four things at once: swapping cryptocurrency wallet addresses on the page, harvesting card details, stealing cookies that lacked HttpOnly and taking screenshots, and loading an exploit against iOS devices.

This is the single sentence from this issue most worth keeping, and it is the research report's own wording:

Cards entered in hosted payment iframes (Stripe, Adyen and similar) were out of its reach. Cards typed into a site's own form fields were not.

In plain terms: if a card number is typed into an embedded payment window such as Stripe's or Adyen's, this malicious code could not reach it; if it is typed into a field your own site built, it could.

Our reading (this paragraph is ours, not the report's): plenty of companies build their own card fields so that checkout "looks a bit better and skips a page", then pass the number on to the payment provider. That decision normally shows no difference at all, and this incident made the difference concrete — when somebody injects JavaScript into your site, an embedded payment window is a wall you do not have to maintain yourself, and a hand-built field is not.

What to do: ask your technical team or your maintenance supplier one question — are our card fields an embedded payment window, or input boxes on our own site? If it is the latter, that does not mean something is wrong today, but it does mean your checkout page has no buffer against "a third-party script got injected". While you are there, ask a second question: how many third-party scripts does our site load right now, and whose are they?

One more thing we only saw before publication: Optimizely had a second incident in the same week, on a different product line. On 19 November the status page opened a separate entry, Update on Configured Commerce incident, and as of our reading the status is Monitoring (last updated 21 November, 00:35 UTC). Verbatim:

Between 10 and 12 November 2026, an unauthorized third party gained access to a legacy Configured Commerce build and deployment server that we were in process of retiring. Optimizely detected the activity on 12 November 2026, removed an injected script (on two sites only), and permanently shut down the compromised server. We have secured our environment and are processing forensics. This incident is unrelated to the Experimentation incident.

The two should be read separately — the company itself stresses that they are unrelated — and this one is far smaller in scale (the company says a script was injected on two sites only). Our reading (this paragraph is ours): from a management point of view they point at the same class of problem — a server "in the process of being retired" still held the right to deploy code to production. Things on the retirement list are often the last to have their access taken away, and until it is taken away it remains an entry point.

2. PHP, 19 November: the highest severity is CRITICAL, but the same paragraph carries a qualifier

PHP published a pre-announcement on 18 November and released four security versions on 19 November: 8.5.13, 8.4.28, 8.3.37 and 8.2.36.

The sentence most often repeated from that pre-announcement is this one (verbatim):

These releases are part of PHP's regular four-week release cycle and will address security vulnerabilities affecting all supported PHP versions. The highest severity is CRITICAL.

And immediately below it sits this one (also verbatim, same announcement):

The issues rated CRITICAL require either non-default configuration or uncommon setups.

These two sentences are two halves of the same paragraph. Quoting only the first manufactures panic; quoting only the second hands people an excuse not to update. We quote both, because the decision in front of you is a different one — it is not "does the whole company work late tonight", it is "schedule this batch promptly through the normal process, and go and confirm one thing: is our configuration the non-default one it mentions".

How many were fixed depends on which branch you are on. We read PHP's official ChangeLog entry by entry (not an aggregator), and counted this:

Your branchUpgrade toFixes in this batch
PHP 8.58.5.1318
PHP 8.48.4.2814
PHP 8.38.3.3713
PHP 8.28.2.3612

Among the 12 present on all four branches, three have one-line official descriptions that say outright "could lead to remote code execution" (timezone cache, partially written session files, inconsistent session serializer settings), and there is also PHP-FPM request smuggling and a local privilege escalation.

Which ones are CRITICAL, PHP itself does not say — its ChangeLog gives one line per entry with no severity column, and the pre-announcement only gave the highest. We took the scores from AWS's bulletin: in its 19 November security bulletin AWS marked seven of them CVSS 9.2. That is AWS's scoring, not PHP's, and we write it that way.

A small number that does reconcile: AWS says this batch addressed 17 issues, while we counted 18 identifiers in the ChangeLog. The difference is one entry whose official note reads only "upgrade uriparser to 1.0.3" — that is the identifier of a third-party library PHP uses, not a PHP issue. 18 minus 1 is 17, and the two sides agree.

One thing AWS states more clearly than PHP does: if you run in the cloud, who is responsible for updating? AWS set it out service by service in the same bulletin — users of Elastic Beanstalk managed platform updates receive the updated platform during their maintenance window; everyone else has to update themselves; and for PHP installed by hand on EC2, ECS, EKS, Fargate or Lightsail, it is entirely the customer's responsibility.

The same page carries a sentence we would very much like you to remember (verbatim): AWS shipped WAF rules this time, and wrote this beside them:

Patching the affected software remains the only complete fix and should be the first priority.

That is AWS's own sentence, not ours: a WAF is not a substitute for patching.

What to do: ask "which PHP version are we on, and when do we upgrade". And a reminder about something that really does happen in practice — installing the update without restarting the service means it has not taken effect. AWS's bulletin also specifically says to restart the services that run PHP, or rebuild and redeploy container images.

3. Laravel published 17 advisories in one day, and the official blog carried none of them

The point of this section is not only the number; it is that we nearly got it wrong.

Laravel's official blog published nothing at all between 16 and 22 November; the most recent post is dated 21 October. Reading the blog alone, we would have written this section as "a quiet week for Laravel".

But the project's own security advisory endpoint on GitHub published 17 advisories within 18 November — by GitHub's severity labels, 15 medium and 2 low, with no high and no critical. They include SQL injection via PostgreSQL query syntax, cross-site scripting in Blade templates, Eloquent mass-assignment bypasses and validation rules being bypassed.

The genuinely useful part of this section is comparing the release dates: Laravel 12.73.0 and 13.43.0 were released on 17 November, with the advisories following a day later — standard coordinated disclosure, fixes out first, explanation afterwards. Looking further back, 6 of those 17 had already been fixed in earlier point releases, and this was simply the advisory catching up.

So if you have been keeping up with Laravel's point releases, then on the day those 17 advisories appeared, 6 of them were already in a fixed state for you. That is what "routine updates on the calendar" is concretely worth, inside a single week.

What to do: ask your team which Laravel version you are on, and whether you track Laravel security news via the blog or via the advisory endpoint. This week the two answers differ a great deal.

4. WordPress 7.1.5: 20 fixes, one of which could lead to remote execution without logging in

WordPress released 7.1.5 on 18 November, stating verbatim: This security release features fixes for multiple issues. and it is recommended that you update your sites immediately. It also says the fixes were backported to all branches still receiving security updates.

The project listed 20 items and gave no severity rating for any of them, but it did write the privilege prerequisite into each description — so "which ones can be attacked without logging in" is not our judgement, it is copied from the official descriptions. Four are marked Unauthenticated:

  • Unauthenticated privilege escalation via Template Part block attributes, which the project says "could lead to remote code execution" — the only one of the 20 where "no login needed" and "remote code execution" appear together
  • Unauthenticated stored XSS via shortcode attributes in comments
  • Disclosure of private taxonomy names via the REST API search endpoint
  • Unauthenticated denial of service via batched XML-RPC pingbacks

Most of the remaining 16 require some logged-in role (Contributor, Author, Editor, Administrator), and two of them are SQL injection.

On wording, we deliberately copy the project's "could lead to". It wrote could lead to, not "this is being exploited", and those are very different statements.

What to do: the WordPress update button is right there in the admin area, and the project itself says to update immediately. If you have automatic background updates enabled, this batch will usually install itself — but that is also something worth confirming rather than assuming.

5. These 37 have one thing in common: not a single CVE number

Adding the last two sections together: 17 from Laravel and 20 from WordPress, 37 security fixes in total, and not one of the 37 carries a CVE number.

A CVE number is the industry's standard identifier for one specific vulnerability. No number does not mean not serious — the "no login needed, could lead to remote execution" item above has no number. The practical consequence is a different one:

If your security process, your scanning tool or your supplier's report tracks vulnerabilities by CVE number, these 37 will be absent as a group. Not rated low risk — simply not on the list.

This week happens to provide a clean control group: every entry in the PHP batch does have a number (and a GitHub GHSA identifier as well), so any tool that tracks by number will raise them. Same week, same kind of software running under your website, and the ones with numbers get seen while the ones without do not.

What to do: this is a question well worth saying out loud — does our vulnerability tracking look at anything besides CVE numbers, such as the advisories each package publishes itself?

6. AWS published 9 bulletins this week, three of them about AI tooling itself

AWS published 9 security bulletins between 16 and 19 November. Besides the PHP one already covered, three of them concern AI-related tools themselves, and the conditions are written out specifically:

  • The AWS pricing MCP server (aws-pricing-mcp-server, fixed in 1.0.31): path traversal and arbitrary file write. AWS wrote one sentence in the bulletin that is well worth reading — the tool was annotated as "read-only", which leads MCP clients to auto-approve it without asking the user again. For cost-analysis-mcp-server in the same batch, AWS states plainly that no fixed version will be published.
  • The MCP server for Timestream for InfluxDB (fixed in 0.0.19): AWS writes that a remote, unauthenticated attacker can use prompt injection to place a forged InfluxDB URL into the MCP client's context. Its mitigation section states verbatim: There is no workaround that fully fixes this issue.
  • Deep Java Library (fixed in 0.39.0): a tampered model archive can read any file the process can access. AWS's mitigation reads verbatim: Only load model archives from sources you trust.

Our reading (this paragraph is ours): none of these three is about "AI being attacked". They are about bugs in the AI tooling itself. The first is especially worth remembering because it is such a clean lesson — a tool labelled "read-only" was auto-approved, and it could in fact write files. The third is the thing we keep saying, with a new object: we used to say "the package was written by somebody else", and now we have to add "the model was packaged by somebody else too".

AWS's OpenJDK bulletin also contains a phrasing worth learning from (verbatim): We have no evidence that this issue has been leveraged at this time. — "we have no evidence it has been exploited" and "it has not been exploited" are not the same sentence, and most retellings turn the first into the second.

7. LiteLLM: the only item this week that CISA says is already being exploited

The United States' Known Exploited Vulnerabilities catalogue (CISA KEV) added 4 entries on 18 November, one of them LiteLLM's CVE-2026-40217 (counting this week's window from 16 November, 9 entries were added in total up to our last reading). LiteLLM is an AI gateway many teams use to reach several model providers through one interface.

The catalogue's description, verbatim in substance: an unprotected alternate channel, through which an attacker can rewrite bytecode via the path /guardrails/test_custom_code to execute arbitrary code. The remediation deadline set for US federal agencies is 2 December 2026.

This is the only item in this issue assessed as "already being exploited". For the PHP, Laravel and WordPress fixes above, none of the vendor announcements mentions any sign of exploitation.

One thing we have to be honest about: everything above about LiteLLM comes from CISA, an aggregating source. We have not yet read LiteLLM's own statement about this identifier. Our own rule is that only the vendor's own announcement can be written as fact, so please treat this section as a signal rather than a complete description.

What to do: if you use LiteLLM, this item outranks everything else in this issue — because it is the only one somebody is already attacking.

8. Debian kept up this time, and incidentally demonstrated that "a fix can go wrong too"

Debian published 12 security advisories between 16 and 19 November, with both php8.4 and nginx on 19 November — the same day as the upstream releases.

That differs from last time in a way worth mentioning: in week 45 we recorded that the nginx batch showed all three Debian releases marked vulnerable with "unfixed" in the fixed-version column. This time the distribution kept up.

And the real teaching material in this section hides in a two-character difference: on 18 November there was DSA-6637-1 needrestart security update, and on 19 November, right behind it, DSA-6637-2 needrestart regression update.

In other words, a security fix needed fixing again the next day.

We have written a piece on whether updates break websites, quoting the UK National Cyber Security Centre's guidance — the official term is "problematic updates", and the remedy it offers is not to stop updating but to stagger updates and have a rollback strategy ready in advance. This week's -1 and -2 are a ready-made example of that sentence.

9. nginx

nginx released 1.31.9 (mainline) and 1.30.7 (stable) on 18 November, handling ten CVEs at once, of which nginx.org rates the HTTP/2 one as major. This topic is covered by the weekly on the Navide side (we divide topics between us to avoid two heavily overlapping weeklies), and we record only versions and dates. Debian's corresponding fix landed on 19 November (see the previous section).

10. What we checked and did not find, and what we did not check

Checked, and not present this week:

  • Symfony: 0 entries in the window. One note here, because the process itself is a lesson — the URL symfony.com/blog/category/security returned 404 for us. That is "the URL could not be read", not "nothing this week", so we moved to the project's own advisory endpoint before drawing the conclusion. November's 21 entries are all dated 4 November, the batch we already covered in the previous issue.
  • Laravel and WordPress have no corresponding Debian security advisories (WordPress's most recent is 29 October).
  • None of the identifiers named in this issue, apart from the LiteLLM one, is in CISA's Known Exploited Vulnerabilities catalogue — and this time that sentence means more than we had expected. The catalogue version we compared against before publication is 2026.11.20, with 1,790 entries, while the PHP batch was released on 19 November: the catalogue is now later than the PHP release, and not one of those 18 identifiers is in it. (When we first wrote this section the catalogue still stood at 18 November, a day earlier than PHP, so "not listed" proved nothing at all then.) It still only says "as of 20 November, CISA has no evidence of active exploitation", which is not the same as "it will not be exploited" — PHP's own pre-announcement wrote that attackers often start looking as soon as a vulnerability is disclosed.

What we did not check (written out, because the half you cannot see is the half you need to know about):

  • WordPress plugins and themes. That section covers WordPress core only. For a small or medium business website, plugins tend to cause trouble more often than core does, and we did not check them at all this issue.
  • For the seven PHP issues AWS marked 9.2, we did not read the corresponding GitHub advisory pages one by one, so "how PHP itself describes the severity" remains blank in this issue.
  • LiteLLM's own statement about CVE-2026-40217 (see section 7).
  • The body of Debian's php8.4 advisory (which CVEs it covers, and the package version it fixes).
  • Sansec, Tenable, Snyk and Checkmarx — these research-side sources were not scheduled this issue.
  • Follow-up on Optimizely's second incident (Configured Commerce): as of our reading the status is still Monitoring, the company says forensics are in progress, and it has not yet said whether any data left the building.
  • GitLab's security advisories this week — we did not check them at all. (We only noticed before publication that we had not even written "we did not check GitLab" — which is worse than missing it, because you cannot see the gap. So it is recorded here.)

The one thing we would like you to take away

More than 70 security issues were fixed this week, and the judgement in front of you really only has three layers:

First layer: is anyone already attacking it? This week there is exactly one, LiteLLM. Second layer: which ones need no login? That WordPress item, and the PHP entries marked remote execution. Third layer: schedule the rest at your regular cadence.

And the question most worth asking out loud today is the one from section 1: are our card fields an embedded payment window, or input boxes on our own site? That question has nothing to do with version numbers and nothing to do with updates. It was settled when the code was written, perhaps three years ago — and this week 178 organisations happened to demonstrate what that decision was worth.

NT IT is a systems software company that grew out of custom development, and what we do now is the long-term maintenance of AI systems. We read these advisories every week because the systems we maintain depend on the information; we write them up because most owners do not have time to follow them. If you are not sure which versions your systems are running, or whether anyone is reading these advisories on your behalf, you can start with a system health check: book a system health check

Want to bring these practices into your own company?

What a Helper CTO does for your system
Book a System Health Check Ask us on LINE

We don't chase volume.

We build long-term relationships with a select few partners worth going deep with.

Need help?

Message us directly on LINE