The breach notification letter is now written by a model, and it has gotten worse
They are longer, vaguer, and warmer. A modest proposal: what was taken, what to do, and whether an AI system had access to the data, in that order.
By Marguerite Okonkwo-Hale, Senior Breach Correspondent4 min read
I have read, by my rough count, four hundred breach notifications in the course of this job, and until recently I could recite one from memory. It valued my trust. It described unauthorized access by a sophisticated actor. It assured me the company takes security seriously. It said "certain information" "may have been" affected, offered twelve months of credit monitoring, and apologized for any inconvenience.
The last fifty have been different, and the difference is that they are longer. A letter that used to run 600 words now runs 1,800. It opens with a paragraph about the company's journey. It has subheadings. It is warm in a way no legal department has ever been warm, and vague in a way no legal department ever quite managed either, because it has replaced "may have been affected" with four paragraphs that circle the same missing fact. Three of the fifty used the identical phrase "we recognize that this news may be unsettling." Nobody recognized anything. A model was asked to draft a notification that was "empathetic, thorough and compliant," and it did.
Who the letter is for
The breach notification is nominally written for the person whose data was taken. In practice it was always written by lawyers for regulators, with the affected person as an incidental reader. Now it is written by a model for lawyers for regulators, and the reader has slipped one further place down the list. The model learned what a notification sounds like from four hundred of them, so it produces the genre perfectly: every obligatory phrase, no obligatory fact.
The vagueness is also no longer entirely strategic. In two recent letters I checked against the company's regulatory filing, the letter and the filing disagreed about which fields were taken. The filing was written by a person who knew. The letter was written by something that did not, and nobody caught it, because the letter read beautifully.
The question nobody answers
There is also a fact the old letter never had to address and the new one is carefully not addressing. Companies have spent two years connecting models and agents to their customer databases, ticketing systems, inboxes and document stores. When one of those companies is breached, I now want to know something I never used to ask: did an AI system have access to my data, and was it part of how the data left?
The answer changes what I should do. A breach through a stolen database tells me to rotate a password. A breach through an assistant that was talked into exporting records tells me the attacker chose what to ask for, may have asked in plain language for exactly the fields useful to defraud me, and may still have the channel. It also says something about the company, which is presumably why the letter does not.
A modest proposal
The notification I want is short and ordered by what the reader needs, not by what the company fears or what a model thinks a letter sounds like.
- What was taken, specifically. Email addresses and hashed passwords. Or names, dates of birth, and government ID numbers. Not "certain personal information." If the hashing was weak, say so.
- What to do, in priority order. Change this password here. If you reused it, change it there. Enable this. Watch for this specific phishing, because the attacker now has the information to write it and will have a model do the writing.
- Whether an AI system had access to the data. Was an assistant, agent, or model-backed feature connected to the affected store? Was it the path in or the path out? What has been done to it since? If no such system was involved, one sentence saying so.
- When. When they got in, when they were removed, when you found out. The gap between those dates is the most useful fact in the letter.
That letter is 300 words. A person could write it in an afternoon. A model could write it in a minute if asked for it, which is the point: the problem was never the tool. The problem is that someone asked for a letter that would satisfy rather than inform, and the tool is now very good at satisfying.
The notification is the one moment a company has the full attention of the people it has failed. It used to spend that moment on itself. Now it spends it on a paragraph about its journey.
The counterargument, addressed
The objection is liability: specificity invites lawsuits. Perhaps. But vagueness invites them too, and it guarantees the reader cannot protect themselves. Regulators in several jurisdictions now require plain-language disclosures with concrete guidance, and the companies that comply do not appear to have been sued into oblivion. A few have been praised, which is a novel outcome for a breach.
So, a rule. Every breach notification should be read, before it is sent, by a person who is not a lawyer and did not use a model to draft it, and who must answer one question: "If my mother received this, would she know what to do?" If the answer is no, it is not finished, however long it is.
About the writer
Senior Breach Correspondent
Marguerite has read more breach notification letters than any living person and has started to dream in the passive voice. She covers incidents, disclosures, and the gap between the two.