Claude Fable got better. Your inbox didn’t.

A hidden prompt is malware for AI — it hijacks your AI instead of your computer.

Thirty years in enterprise software taught me one pattern: we adopt the new thing faster than we secure it. Right now the new thing is AI doing real work — reading your email, summarizing your documents, taking action on your behalf.

Here's the problem nobody tells you about. An attacker can hide instructions inside an ordinary email or attachment — invisible to you, perfectly readable to your AI. We've always had a word for code that hijacks software: malware. A hidden prompt is malware for AI. It just hijacks your AI instead of your computer.

The AI companies are working on it, and the best are getting good. Claude Fable 5 — the model most resistant to these tricks — still gets fooled about five times in a hundred when an attacker keeps trying.

And anyone with your AI agent's email address gets unlimited tries.

My co-founder Oleksii built his career protecting military networks with one rule: trust nothing, rebuild everything. That's InZero. Every email, message, and file is taken apart in a sealed room and rebuilt as a fresh copy that looks exactly like the original. Hidden instructions, viruses, invisible text — gone, because they never get copied over. The original is locked away as evidence, your security team gets a report, and it sets up on Microsoft 365 in minutes.

One more thing. A rebuilt copy isn't just safe to read — it's safe to remember. The same cleaning can fill your company's AI memory with only clean, trusted information.

Clean every email, chat, and file. Protect your people and your AI. Fill your company brain only with what matters. That's the company.

— Karl