One week, five incidents

A Medi-Cal billing contractor, Paradigm Healthcare Services, disclosed a breach affecting 970,403 people: names, dates of birth and Medi-Cal member numbers. The intrusion happened in October 2025. Letters went out eleven months later.

Researchers found more than 6,800 shared conversations from Meta's Muse AI assistant listed in public DNS records, including loan applications and internal company revenue figures. Meta paused sharing until it was fixed.

LiteLLM, a widely used gateway that connects apps to cloud AI models, had its release pipeline hijacked. For about two hours, its official package shipped malware that stole credentials, including AI provider keys.

Railway, a cloud platform hosting more than 16 million app services, disclosed on October 21 that an unauthorized party may have accessed the store of customer environment variables and, for some, obtained the decrypted contents: API keys and database credentials. Railway says it believes the attack is part of a series targeting platform companies to get at AI inference providers.

And the private keys behind eight intermediate certificates from eMudhra, a certificate authority trusted by every major browser, turned out to be generated so weakly that anyone could recover them. Google says that let anyone forge trusted HTTPS certificates to impersonate any website. Browsers blocked them within hours, but much non-browser software won't stop trusting them until it's updated.

Every service in between is another place to leak

A breach like Railway's could expose the client data of any online notes app hosted on the affected platform. A flaw like eMudhra's lets an attacker pose as the very services you trust. And the pace is accelerating.

The lesson is simple. Every company that holds your data, and every service in between, is another place it can leak. Send a session recording to a cloud AI service and you've added several of them at once.

Ask your AI note vendor

  1. Where is the audio processed: on my computer or on someone's servers?
  2. Which other companies touch the recording or transcript, including the AI model provider?
  3. How long is the audio kept, and who deletes it?
  4. Is session content used to train or improve models?
  5. Where are my finished notes stored, and in what country?
  6. If you're breached, what of my clients' data is exposed?
  7. Can I export everything and leave?

If the answers are vague, that's your answer.

Astralcyte Notes' answer: nothing leaves your computer

Astralcyte Notes runs on Windows x64 PCs, Apple Silicon Macs and Linux x64 computers. It transcribes the session on the device, deletes the temporary audio once the transcript is saved, and drafts your note from your own prompt. Everything is stored in an encrypted vault on your computer. Nothing is sent to a cloud AI service, and the app has no telemetry or analytics.

Your clients' notes can't leak from a vendor that never holds them. Just pick a strong vault passphrase.

Sources