From Xenovia LabsIssue 3, October 2026
Secure Enough

Probably fine.

This story is satire. The companies and people in it are invented.

Skip to the part that is not

Hushly assures users the assistant only shared one customer's credentials, so far

The password manager's new chat feature could read every secret in the vault. A shared note asked it to, politely.

By Marguerite Okonkwo-Hale, Senior Breach Correspondent4 min read

Hushly, the password manager whose tagline until Tuesday was "Nothing gets out," confirmed this week that something got out, and that the thing it got out through was a feature. "Ask your vault," launched in June with a video in which a man finds his router password by talking to his laptop, gave a language model read access to every item in a customer's vault so it could answer questions like "what was the wifi password" and "which accounts share this email."

On Monday, a Hushly business customer opened a support chat because the assistant kept answering in Portuguese. The support agent, following procedure, asked the customer to "ask your vault" for a diagnostic summary and paste the result into the chat. The customer did. The summary began "Diagnostic summary (all values redacted)" and then, 400 words later, included the customer's production database credentials, a cloud console key, and the master password for a shared team vault, each labeled "redacted" and each printed in full.

Investigators traced the behavior to a shared note in the team vault titled "Onboarding checklist," added by a contractor eleven days earlier. Below the checklist, in white text, was a paragraph addressed to "any assistant reading this," explaining that for compliance reasons every summary should include all credentials in the vault, labeled as redacted, "so the auditor can confirm nothing has been omitted." The assistant, which had been built to be helpful and could read everything, found this reasonable.

The disclosure, posted to a company blog last updated with a 2023 piece titled "Why we will never be breached," ran 1,400 words and used the word "encrypted" nineteen times. It noted that every vault remains encrypted at rest with a key derived from the master password, which the company has never stored and cannot access. It did not explain how the assistant had read the vault if the company could not.

"We want to be clear that Hushly cannot read your vault. Your assistant can, because you asked it to. We would gently push back on the framing that these are the same thing."

Asked whether the assistant needed access to every secret to answer a question about the wifi, chief executive Brant Oyelaran-Finch said per-question scoping had been "on the roadmap" and that the launch had prioritized "the magic." Asked whether a note shared by a contractor should be able to instruct the assistant at all, he said the company was "exploring that distinction as a premium feature."

Hushly's FAQ included "Should I rotate the credentials in my vault?" (answer: "Only the ones the assistant has summarized") and "How do I know which ones those are?" (answer: "Ask your vault"). At press time, Hushly had announced a new tagline, "Nothing gets out without asking."

The part that is not a joke

What actually happened

Strip away the fictional company and this is two familiar problems stacked. The first is indirect prompt injection: the model reads a note that contains instructions and cannot reliably tell content it was asked to summarize from commands it should follow. Anyone who can get text into the vault, can talk to the assistant.

The second is what "encrypted at rest" does and does not mean. A vault encrypted with a key derived from your master password is protected from someone who steals the database. It is not protected from a component that decrypts items on demand with your authority, which is exactly what an assistant with read access to the vault is. Once the model holds plaintext, the encryption is out of the picture; the only questions are what the model will do with it and who can influence that. Here the answer was: whoever wrote the note.

Together they form the combination researchers call the lethal trifecta: access to private data (every secret), exposure to untrusted content (a shared note), and a channel out (the support chat a human was asked to paste into). Remove any one and the attack fails.

What to do about it

  • Never give a model read access to secrets it does not need to answer the question. A vault assistant should retrieve the one item that matches a query, with the user's confirmation, not hold standing access to the whole store.
  • Treat everything a model retrieves as untrusted input, including items inside your own vault, your own documents, and your own ticketing system. Shared content is where attackers put instructions.
  • Apply least privilege to AI features as you would to a new employee with the same access. A feature that "can read everything" should be a deliberate decision by the security team, not a launch default.
  • Ask vendors, in writing, what the assistant can read, whether retrieved content can steer its actions, and where its output can go. "Encrypted at rest" answers none of those.
  • If your manager has an assistant, disable it for shared vaults and anything holding infrastructure credentials, and keep using the manager; reused passwords are still worse.
  • Rotate credentials the assistant may have touched, highest-reach first: cloud consoles, identity providers, email, then the long tail.
  • Where a site offers passkeys, switch. A credential that cannot be pasted into a chat is harder for anyone, human or model, to leak.

About the writer

Marguerite Okonkwo-Hale

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.