Watson Standard1
Research · outside companies Written by Bryce Watson

← All research

You can be hacked through your host, a script you load or a package you install. Here are seven attacks from the past five weeks.

If a bigger customer sends you security questions, some may be about the companies your product relies on. I keep a weekly log of security news, with AI agents I run doing the reading. On November 17, 2026 I had them sort the 42 incidents and flaws that got their own entry in a month of that log and its notes, by how each one got in. Twenty-three were flaws that an update fixes. Nine came in through an outside company or a stolen sign-in, and in those the affected product's own code could have been fine.

The short version

Write down every outside company and account that can reach your product, and for each one how you'd cut it off and replace its keys. Section 3 has six checks you can run this week without hiring anyone.

Published . Sources read November 17 and checked again November 21.

1 What the log holds

The log covers October 16 to November 17, 2026. Each week it takes in the US cybersecurity agency's list of flaws under real attack, GitHub's reviewed advisories for JavaScript, Python and PHP, a set of news searches and a weekly news recap. Each incident or flaw with its own entry got one row, 42 in all measured, and each row was sorted by the way the problem reached the product derived.

How it got inRowsWhat the owner could do beforehand
A flaw in software the product runs23 of 42Install the update, or ask the host to
A company the product runs on, or one that stores its keys or its owner's data, was broken into6 of 42Know which companies those are, and be ready to replace keys
Outside code tampered with before it arrived, or pushed in through a stolen sign-in3 of 42Limit what that code can touch
A developer's own tools or computer7 of 42Update the tools, and take care with what they open
The owner's own settings3 of 42Fix the setting

The second and third rows are the nine derived. Seven of them are the incidents in section 2. The other two are a second web host and a certificate company, which I know only from news reports, so I count them and don't describe them. Updates were still the largest group in this log, and installing them wouldn't have stopped any of these nine.

This counts what reached my reading list. It doesn't count everything that happened. Two of those sources, the agency's list and GitHub's advisories, only list flaws in published software, so a mistake inside one company's private code never shows up in them. It leans toward the languages and tools I watch. And the sorting is my judgment. A stricter reader could drop two of the nine, one of which reached no product at all, and count seven derived.

2 Five ways in, from seven incidents

AI agents I run opened each of these at the affected company's own report or the researcher's own write-up on November 17, 2026, and again on November 21. The numbers and quotations are the sources'.

A script another company serves on your pages

Optimizely makes a testing tool that its customers load on their own web pages. On November 16, Report URI fetched 1,611 of those scripts and says 188 carried malicious code, belonging to 178 organizations, and that "the real number is likely higher" because it could only check the ones it found in time. Its list includes some of the largest brands on the web. It says the code swapped cryptocurrency addresses, copied card details typed into forms, and sent cookies and screenshots out. Report URI saw every infected file it had found cleaned by the end of that day, and says it doesn't know how the attacker got in reported (Report URI). Optimizely's status page said on November 17 that an unauthorized party "exploited a vulnerability" in the product to insert code into some customer projects, and that it had fixed it. On November 18 it said the flaw was in the product's account signup flow, that "Less than 5% of our customers were impacted", and that it had contacted every customer of that product directly reported (Optimizely).

The code reached those sites' visitors through a script the sites load from Optimizely.

A neighbor on a shared server

Hostinger, a web host, says an attacker "first entered on November 5 through a vulnerable content management system on a customer website", then used a flaw nobody knew about in the web server software to get past the wall between accounts and take over the shared server. Hostinger says the attacker "ran code or read credentials on about 400 accounts" on that one server, and that it emailed affected customers individually. It runs more than 5,000 servers. When read on November 21, the report still said the investigation was ongoing reported (Hostinger).

A service that stores your keys

Apps keep their database passwords and API keys in the settings of the services that run them. Trigger.dev, which runs background jobs, says it learned on November 1 that an attacker had obtained keys that give "full access" to the settings of some customers' environments, that this "impacted 0.66% of all customers", and that those customers should treat every credential stored there as compromised. It hasn't said how the attacker got in reported (Trigger.dev).

It was the third such notice in the log in three weeks. Zeabur, a hosting platform, confirmed that 3,218,965 stored settings records were read on October 15 reported (Zeabur). Railway, another, said on October 21 that stored settings may have been read, decrypted for some, and strongly recommended, as a precaution, that users replace all of them immediately. Its fuller report of November 10 says the attacker obtained and decrypted stored settings for at most 1.85% of workspaces, about 100,000 reported (Railway, Railway's report).

A package update

LiteLLM is a Python library for calling AI models. On October 21 a release of it, version 1.107.0, went out carrying code that ran every time Python started and sent the machine's credentials away. Datadog says it was available for "approximately two hours" before it was pulled, that LiteLLM's Docker images tagged latest carried it too, and, citing LiteLLM's own analysis, that the attackers used sign-in tokens stolen from three engineers' accounts to run the project's release process reported (Datadog). Socket, which flagged the release seven minutes after it appeared, says its evidence doesn't establish how a key step of the release was authorized reported (Socket).

An app that installed the newest release during those two hours, or pulled the Docker image by a tag like latest, got it. One that installs from a lock file at deploy and pins its images to a fixed version didn't, unless a person or an update bot moved the version in that window. The same goes for a developer's computer or a build job: any machine that installed it in that window should be treated as exposed derived.

A stolen sign-in to your code

Aikido, a security firm, says that starting November 10, people using 4,837 compromised GitHub accounts opened a probe pull request in 61,764 repositories, and in some of those then pushed a script that sends out the secrets a repository's automated jobs hold. It counts 10,431 of the probes in repositories the account didn't own, reached through shared access, and at least 1,835 repositories where the script ran and credentials were taken. It calls its counts "based on data we've currently collected" and doesn't say how the accounts were taken over reported (Aikido).

So a repository could be reached through its owner's stolen sign-in, or through that of anyone else with access to it, such as a contractor derived.

3 What you can check this week

You can do these without hiring anyone. They won't catch everything. Together they give you one list of the outside companies and accounts that can reach your product.

  1. Scripts from other companies. Open your site in Chrome or Edge, press F12 (Cmd+Option+I on a Mac), choose the Network tab, click the JS filter and reload the page. Right-click the column headings and switch on Domain. A row whose domain isn't yours is a script another company serves. Do the same signed in, and on your sign-in and payment pages. One page load only shows what loaded that time, so if you use a tag manager, open it and list what it loads too. For each script, write down who added it and whether it's still used. Before anyone removes one, check the sign-in and payment pages still work without it.
  2. Services that store your keys. Open the settings page of your host and of any background-job, build or automation service, and find where it keeps environment variables or secrets. Write down each key's name, never its value, and next to it the service that issued the key, since that's where you'd replace it.
  3. Notices you may have missed. Hostinger and Trigger.dev both say they told affected customers by email reported (Hostinger, Trigger.dev). For each company on your list, check which email address it writes to, since it may be a former developer's. Then search that inbox and its spam folder for "security incident" and "security notice", and look at the company's status page or blog.
  4. Who can change your code. In your GitHub repository, open Settings, then "Collaborators & teams", and remove anyone who no longer needs access. Then open Settings, "Secrets and variables", "Actions", and for each secret listed, ask what it opens measured (GitHub, GitHub). Your own account counts too: review the apps and access tokens it has authorized. Secrets can also be set for one environment, under Settings, "Environments", and for a whole organization in the organization's settings. Deploy keys and installed GitHub Apps can change code too, and they have their own pages in the same Settings. Aikido's write-up lists the signs of the November campaign, including a pull request titled "docs: dependency audit notes", and says that if the script ran, every secret and cloud role the repository's automated jobs could reach should be treated as compromised reported (Aikido).
  5. Fixed versions. Look in your project for a lock file, such as package-lock.json, pnpm-lock.yaml, uv.lock or poetry.lock, and check that it's saved with your code. It records the exact version of every package. It only protects you if your deploy installs from it (npm ci and uv sync --locked do, for example), so ask whoever set up your deploy. If you run software from a Docker image, ask whether the image is pinned to a known-good version, or better its digest, and not to latest.
  6. Shared hosting. Ask your host one question: "Is my site on a server shared with other customers' sites?" If it is, find out now how you'd change every password and key stored in your site's files, starting with the database password, and how you'd restore the site's files from a clean copy, so you can do both the day your host writes to you.

Keep the list. The Cloud Security Alliance's Cloud Controls Matrix, the basis of its questionnaire for assessing cloud providers, has a domain named "Supply Chain Mgmt, Transparency, & Accountability" measured (Cloud Security Alliance). This list is where your answers to questions like those start.

4 What I checked and changed on my own side

AI agents I run did the searching below. Each change went in on my word.

How I got this

The log is one I keep for my own work: six passes and two shorter checks between October 26 and November 17, 2026, the first pass reaching back to October 16. Most of the passes were the weekly scheduled run. AI agents read the sources and wrote the rows. One of the 42 comes from the notes kept beside the log. On November 17 an agent sorted the rows by route into the five groups in section 1, a second agent challenged the sorting, and three rows moved to other groups as a result, leaving the nine. Lists the log only names in passing, such as other entries on the agency's list or batches of poisoned packages, aren't in the 42.

I sell security reviews, which is a reason to check this against the sources below. The labels say how far each claim goes:

Measured
I or an AI agent I run counted it, or read it on the source that owns the fact (a product's own documentation, say).
Reported
A named source says it, and I can't check it myself.
Derived
My own reasoning or arithmetic from the two above.

Sources

All read on November 17, 2026 and checked again on November 21, except Railway's November 10 report, first read on November 21.

  1. Report URI, Scott Helme, "World's Largest Brands Infected With Malware Via Optimizely", November 16, 2026, updated November 17. blog.report-uri.com
  2. Optimizely status page, security incident updates of November 17 and 18, 2026. status.optimizely.com
  3. Hostinger, its report on the November 2026 shared server incident, November 9, 2026, updated November 10. hostinger.com
  4. Trigger.dev, security incident report, November 6, 2026. trigger.dev
  5. Zeabur, October 2026 security incident report, October 19, 2026. zeabur.com
  6. Railway, security disclosure, October 21, 2026. blog.railway.com
  7. Railway, "Incident Report: October 21st Security Attack", November 10, 2026. blog.railway.com
  8. Datadog Security Labs, its analysis of the LiteLLM 1.107.0 release, October 21, 2026. securitylabs.datadoghq.com
  9. Socket, its analysis of the LiteLLM 1.107.0 release, October 21, 2026. socket.dev
  10. Aikido, its report on compromised GitHub accounts pushing a secret stealer, November 12, 2026. aikido.dev
  11. GitHub documentation, "Managing teams and people with access to your repository" (undated). docs.github.com
  12. GitHub documentation, "Using secrets in GitHub Actions" (undated). docs.github.com
  13. Cloud Security Alliance, "Cloud Controls Matrix" (undated). cloudsecurityalliance.org

Want me to look at your product?

Have a customer's security questions, or a public repo? Ask for a free written read. Once I've said I can take it, you send them, and I send back in writing the three things most likely to stall the deal. It's free. I read what you send, never your live app, and I never use a key I find.

Free written read

Or check it yourself first: the security self-check walks through eleven checks, most of them in your database's dashboard or your code repository (its examples use Supabase and GitHub).