Insights

AI Security

Writing the customer notice when you can't confirm what was viewed

Anthropic's Nov. 7 notice to business customers said an outside firm accessed account and contact data, that it could not confirm whether any particular person's information was viewed, and left the decision to notify staff with each customer. HealthTech vendors should have that notice drafted before they need it.

November 10, 2026 · 6 min read

Writing the customer notice when you can't confirm what was viewed

A breach notice that says "we cannot confirm whether any particular person's information was viewed" moves the hardest decision in the incident from the vendor to the customer. That is what Anthropic's Nov. 7 email to business customers did, according to the people who received it and published it.

I have not found a public statement from Anthropic about the notice, and Daily Byte reports that Anthropic has not published one. Everything below about its contents comes from reporting on the email itself, which security researcher Jamieson O'Reilly posted on X. I am using it because it is a clean, public example of a notice written under uncertainty, and HealthTech vendors who process health system data will write one like it someday.

What the notice said, according to reports

The email was titled "Notice Regarding Your Organization's Account Information," according to Chip Dispatch. It said an outside cybersecurity firm had accessed data without approval while "conducting a security testing exercise on its own initiative." Anthropic did not name the firm.

For enterprise Claude accounts, the listed fields were account ID, organization ID, plan tier, usage and spend metrics, plus business contact details. The notice excluded conversation content, passwords, payment card information and bank account information. Anthropic said it disabled the firm's access, worked to fix the vulnerabilities and opened an investigation. It also quoted a written certification: "The cybersecurity firm has certified in writing that it has deleted the accessed data and will not use it."

Then the two sentences that matter for this post. The notice said Anthropic "cannot confirm whether any particular person's information was viewed," and that it was writing "out of an abundance of caution should you wish to notify individuals in your organization." The decision to tell staff sat with each customer.

Seoul Economic Daily reports that Anthropic reported the incident to South Korea's Personal Information Protection Commission, which is reviewing it, and that about 100 Korean enterprise accounts were involved. The notice itself gave no count of organizations or people, according to Chip Dispatch.

Why the uncertainty sentence lands on the customer

Read the notice from the receiving side. A security or privacy lead at the customer now has a list of field types, no count, no dates, no confirmation of what was seen, and a decision to make about their own people. O'Reilly's follow-up, as Daily Byte summarizes it, pressed on exactly that tension: how a company can say it cannot confirm whether any individual's data was viewed while also telling recipients they face no added risk.

I do not think "we can't confirm" is a bad sentence. Often it is the honest one. Access logs may show that a system was reached without showing which records were rendered. The problem is when that sentence arrives alone. The customer cannot reason about scope from a statement of what you don't know. They need what you do know, stated as precisely as you can, and a date when you will know more.

Now move it to healthcare. A HealthTech vendor holds health system contacts, support tickets, call transcripts, maybe patient-adjacent metadata. The customer receiving your notice has its own obligations, its own privacy office and sometimes its own regulators. If your notice leaves them guessing, their next email to you will be a long questionnaire with a short deadline.

A notice runbook for "we can't confirm what was viewed"

Write this before an incident, as a template your incident lead and counsel have already reviewed. Fill it in during the incident. I would build it in seven blocks.

  1. What happened, in two sentences. Who accessed what class of system, whether it was authorized, and the date range. If you don't know the start date, say what your logs cover and from when.
  2. Data classes in scope, per customer. Not a global list. Tell each customer which of their fields lived in the affected system. Anthropic's notice split enterprise fields from general contact fields, which is the right instinct. Go further and say which tables or objects.
  3. What you can confirm and what you can't. Separate three states: data that was in the reachable system, data your logs show was queried or exported, and data you can't rule in or out. "We cannot confirm" belongs only in the third bucket.
  4. Why you can't confirm it. One sentence on the evidence gap, such as "the system logged authentication but not record-level reads." That tells the customer whether better answers are coming or will never exist.
  5. What you did. Containment, credential rotation, the vulnerability fix, and any assurances from the other party. If the assurance is a third party's written certification of deletion, call it that. It is a promise, not a forensic finding.
  6. What you recommend the customer do. If you leave notification to them, give them material to make the call: a suggested staff notice, phishing guidance tied to the exposed fields, and a named contact who can answer follow-up questions within a stated time.
  7. When the next update comes. A date, even if the update is "no new findings."

Block 6 is where most notices stop short. A vendor that says "you may wish to notify individuals" and attaches nothing has handed the customer work without tools.

The healthcare additions

Health systems will read your notice against their own contract and their own rules, so a few more items belong in the template.

Map each data class to the contract. If any field in scope falls under a business associate agreement, the notice should say so and point to the notification terms in that agreement. Do not leave the customer to discover that from their own records.

Name the regulators you have told. Anthropic's notice reportedly said it reported to "the relevant Korean authorities" without explaining why, according to Chip Dispatch. A customer will ask whether that choice says something about where data sat or whose data it was. Explain the reason in one line.

Keep the third party out of it unless you can confirm. If an outside party was involved and has not been formally identified, describe its role, not its name. Speculation in a customer notice becomes a quote in someone else's article.

Fix the logging gap that forced the sentence. If record-level read logging was missing on a system holding customer data, that is a finding. Put it in the remediation section with a date.

Test it before you need it

Take this template into a tabletop. Give the incident lead a scenario where an outside party reached an internal system with customer contacts and support data, and the logs show access but not reads. Ask them to produce a draft notice for one named customer within the exercise window. Then hand that draft to someone playing the customer's privacy officer and see which questions come back.

If the first question back is "which of our people?", decide now whether you could ever answer it from the logs you keep today.