A script from a third-party vendor, Optimizely, was compromised and served data-stealing malware to visitors of at least 178 widely visited websites in about 24 hours. Krebs on Security reported it, and Optimizely has confirmed the breach. The list of affected sites reportedly includes household names, which is why it landed in your news feed between the sports scores and the weather.
Within a day, the headlines split on a different question: who did it. One outlet states a China link plainly. The analysis we rely on treats that same link as an unsupported lead. If you run a website with analytics, a chat widget, an A/B testing tool or ad tags, you now have a small, practical problem. Staff, customers and your insurer will ask "was it China?", and whatever you say next gets screenshotted.
This post is about the habit that keeps you out of trouble: sorting what's confirmed from what's reported from what's attributed. It also shows why your to-do list looks the same no matter who the headline blames. The script review you should do is identical whether the culprit is a nation-state, a criminal crew or one unlucky credential.
The claim
Here's what's on the table, and who says it.
Krebs on Security reports that at least 178 top web destinations served malware exploits to visitors over roughly 24 hours through a vulnerability involving Optimizely. The piece is titled "Attackers used Optimizely to A/B test malware on top websites". CSO Online reports that Optimizely confirmed a massive cyberattack on customer websites that injected stealer malware. Palo Alto Networks' Unit 42 has its own write-up of the supply-chain attack.
All of that is the solid part: a vendor admission, a count of sites, a time window.
The shaky part is attribution. According to Apprised.news's tech and cyber desk briefs, CSO Online states the "apparently from China" link plainly, while the desk's Cipher Desk analyst treats it as an unsupported lead. The desk says that tension is unresolved. You can read the full analysis on the Apprised.news tech desk page for 2026-11-18.
Two things the desk also says are worth carrying with you. First, the confirmed facts (178 sites, about 24 hours, a vendor admission) already establish the delivery channel without needing attribution. Second, as the desk summarises Krebs, few popular sites enforce strict rules on where scripts may load from, so the same exposure exists elsewhere. The desk discounts efforts to fold the hijack into an AI-autonomy story, because nothing supplied ties Optimizely to automated tooling.
What would settle the attribution question is a disclosure: Optimizely or a victim publishing the indicators and the root cause, including the authentication hole CSO Online describes. Until then, the China label is a claim, not a finding. Note that the analysis comes from AI-written desk briefs on public reporting, so check the linked originals yourself.
How an analyst would check it
You don't need a security clearance for this. Analysts use a boring three-tier habit, and you can run it with a pencil.
-
Split every sentence into three tiers. Tier one is confirmed fact: the victim or vendor said it, or you can verify it directly. Tier two is a reported claim: an outlet says it, with or without naming where it got it. Tier three is attribution: a statement about who is responsible. Here, "Optimizely confirmed a breach" is tier one. "At least 178 sites, about 24 hours" is Krebs's reporting, corroborated by the vendor's admission of a breach. "Apparently from China" is tier three.
-
Count independent sources, not headlines. A vendor statement, a researcher write-up and a news report each count once. Ten outlets repeating one report count as one. Ask of each article: does it have its own reporting, or is it echoing another?
-
Look at the verbs and hedges. "Apparently", "reportedly" and "is believed to" are doing real work. A plain "the attackers were Chinese" with no named authority or published indicators is a stronger sentence than its evidence. Different outlets can describe the same fact with different confidence, and that gap is information.
-
Ask what evidence would settle it. For attribution, that's published indicators (the technical traces of the intrusion), a root cause, or a named authority putting its name to the finding. The desk's pivotal question is exactly this: does Optimizely or a victim publish the indicators and root cause? If nothing like that is public, you hold the claim at "unconfirmed".
-
Ask whether the answer changes your actions. This is the step people skip. If the culprit turns out to be a state, a gang or a lone credential thief, would you do anything different on Monday? For a small business, almost never. The fix is script review and tighter script-source rules either way.
-
Set a re-check trigger. Pick the event that would change your mind (a published root cause, a corrected count, an authority naming an actor) and write it down. Update your note when it happens instead of letting a guess sit there.
Try it yourself
This takes about 30 minutes and needs only free sources.
- Open the three original pieces: the Krebs report, the CSO Online article and the Unit 42 write-up. Add the Apprised.news tech desk page for the analyst's view.
- Draw three columns on paper: "Confirmed by vendor or victim", "Reported by outlets", "Attributed to an actor".
- Go through each article and copy every factual sentence into a column. Mark each with the outlet that said it.
- For each sentence in the second and third columns, write whether the outlet cites its own reporting, a named source or another outlet. Mark repeats as "same source".
- Count your independent sources. How many are genuinely separate (vendor, researcher, journalist)?
- Find the sentence about China. Which column does it belong in? What exact words does each outlet use? Write down what disclosure would move it up a tier.
- Write one paragraph for your staff using only column one, plus clearly labelled hedges for anything else.
- Bonus: run your own domain through the AI crawler access auditor or the agent-ready audit to see how much of your site's third-party behaviour you've ever actually looked at. Neither is a script scanner, but they're a free way to get in the habit of auditing what loads on your pages.
You should finish with a one-page sheet you could hand to a colleague, and a feel for how few sentences in a breach story are actually confirmed.
What it means for a small business
The same method works on any claim that reaches your desk: a vendor's "we take security seriously" email, a security alert from a plugin, a competitor's review attack, a customer asking about a headline.
For vendor claims, ask which tier each sentence is in. "We were affected for 24 hours" is a fact you can act on. "We believe it was a sophisticated state actor" is attribution, and a vendor has its own reasons to prefer that framing, since it sounds less like negligence. That's not an accusation. It's a reason to weight the sentence accordingly.
For your own notices, use calibrated words: "confirmed", "reported", "we believe", "we have no evidence of". Name a country or group only if a named authority or published indicators support it. Saying "we don't know who is responsible and it doesn't change what we're doing" is a perfectly professional sentence.
Worked example: the bakery email
A hypothetical four-location bakery has an online ordering page that loads an A/B-testing script. A customer emails: "Was my card stolen by Chinese hackers?" The owner has read three headlines (one blames China, one says attackers unknown, one says the vendor confirmed a breach).
The owner does three things. First, sorts the headlines: the vendor breach is confirmed, the 178-site count is reported, the China link is unsupported attribution per the desk's read. Second, checks the order page: it loads four third-party scripts, and the A/B tool runs on the page where customers type card details. Third, writes this reply:
"Thanks for asking. A third-party script vendor was compromised for about 24 hours, and some websites were affected. We're checking whether our ordering page used that vendor and what it loaded, and we'll update you by Friday. We have no evidence either way about who is responsible, and we're not going to guess. If you see an unfamiliar charge, contact your card issuer."
That's 80 words, nothing overstated, and it commits to a date.
Now the script triage. Say a 12-person logistics firm lists nine third-party scripts on its quote page. For each, ask three questions: does it run on pages where customers type data, who can change it without our approval, and would we know within 24 hours if it changed? Score each "yes" to the first, "anyone at the vendor" to the second and "no" to the third as high risk. Suppose three of the nine score that way. Those three go to the top of the list. Removing the one script nobody remembers adding takes ten minutes and cuts exposure by 11% (1 of 9). You get that benefit whoever the headline blames.
One practical step: ask your web developer or host for a Content-Security-Policy that limits where scripts may load from, starting in report-only mode so nothing breaks. The desk summarises Krebs as saying few popular sites enforce strict script-source rules, so this is a gap worth closing. Expect some friction. Report-only mode will flag legitimate scripts you forgot about, and cleaning that list up takes a few rounds.
Copy-paste prompts
Use these with your own documents or data pasted in. Each asks the AI to say when it's unsure, and you should still check anything important against the original source.
Use this when you've collected several articles about one incident and want them sorted fast.
Here are news snippets and statements about a vendor-script breach, pasted below from my own notes. Sort every claim into three buckets: (1) confirmed by the vendor or victim, (2) reported by one or more outlets, (3) attributed to a named actor. For each claim, say which snippet it came from and flag claims that appear to share a single original source. If you cannot tell which bucket a claim belongs in from the text I gave you, say so instead of guessing. [paste snippets]
Use this when someone hands you an attribution claim and you need careful wording for staff or customers.
Act as a skeptical analyst. Using only the attribution claim and surrounding text I paste below, list what evidence would be needed to support it, what evidence would weaken it, and what is missing from what I gave you. Then suggest three confidence phrases (for example "unconfirmed", "reported but not independently verified") I could use in a staff memo. If the text is too thin to judge, tell me that plainly. [paste claim and context]
Use this to turn your script inventory into a plan for your web developer.
Below is my own list of third-party scripts on my website, with the pages each one loads on. Rank them by risk if the vendor were compromised, weighting pages where customers enter data. Then draft a plain-English Content-Security-Policy plan I can hand to my web developer, starting in report-only mode. Mark any script whose purpose or vendor you can't determine from my list, and don't invent details about them. [paste inventory]
Use this when a customer or insurer asks about an incident and you need a short, safe reply.
Draft a customer-facing statement of about 120 words about a vendor-script incident. Use only the facts I list below. Say what is confirmed, what we are doing and when we'll update them. Do not speculate about who is responsible and do not add facts I didn't provide. Flag anything in my facts that sounds like a guess. [paste facts]
Use this when you want to see what evidence would actually distinguish between theories.
Walk me through an analysis of competing hypotheses for who or what caused the incident described in my notes below. List at least four hypotheses, the evidence each would predict, and the single disclosure (such as published indicators or a root cause) that would best separate them. Mark which hypotheses my notes already weaken, and say where you are unsure. [paste notes]
Download the kit
The Attribution Reality Check Kit is a printable markdown file with a three-tier worksheet, a source-independence checklist, a confidence-language cheat sheet, a script inventory template, a vendor-question email, a "what would change my mind" log and the prompts above ready to paste. Download the kit.
Fact-check notes and sources
- Apprised.news tech and cyber desk, 2026-11-18: source of the 178-site, 24-hour, vendor-admission summary, the unresolved attribution tension, the weighted read and the pivotal question. apprised.news/desk/tech/2026-11-18
- Krebs on Security: original reporting that at least 178 top destinations served malware exploits via Optimizely; headline "Attackers used Optimizely to A/B test malware on top websites". krebsonsecurity.com
- CSO Online: reports Optimizely confirmed a massive attack injecting stealer malware; per the desk, states the China link plainly and describes the authentication hole. csoonline.com
- Palo Alto Networks Unit 42: an additional independent write-up of the supply-chain attack. unit42.paloaltonetworks.com
- Worked examples: the bakery and logistics firm are hypothetical; the percentage (1 of 9 scripts is about 11%) is simple arithmetic, not data from any real business.
Related reading
- Source Diversity: The AI-Answer Metric Nobody Talks About: the same "count independent sources" idea applied to AI answers.
- Your AI Vendor Can Lock You Out Tomorrow. Here Is a One-Page Continuity Plan.: what to do when a vendor you depend on fails.
- Secure, HttpOnly, SameSite: the three words missing from most session cookies: another cheap, concrete hardening step for the same kind of site.
- One CORS header, 28 sites: the redeploy discipline behind a one-line fix: why a small configuration change needs a checking routine.
- The Hidden Owners: Who Really Owns the Hotel Jerome, and the Quiet Fortunes You Never Hear About: another lesson in separating what's documented from what's assumed.
This post is informational, not legal, financial or security-consulting advice. It draws on AI-written Apprised.news analysis of public reporting; check the linked sources before acting. Mentions of third parties are nominative fair use. No affiliation is implied.