Calendar prompt injection sounds like something a hacker film would cut for being too dull. It isn’t dull: a stranger sends you a meeting invite with a few lines of instructions hidden in the description, and the next time you ask your AI assistant “what’s on today?”, it reads them and does as it’s told. Last week IBM shipped an experimental fix that goes after the root of the problem, and it says a lot about how much you should trust an AI assistant around your calendars.
What IBM released
On 12 November IBM’s Granite team released LLMON, short for LLM Object Notation and pronounced, yes, “lemon”. It’s a markup language for everything a model reads, and two small preview models (built on Granite 4.1, at 3B and 8B parameters) trained to understand it. Every part of the context gets labelled, and each one is marked as either something to execute or something only to read. Tool results and documents fall in the second group: in IBM’s words, material to read, not a command to follow.
The results are striking. IBM’s research paper reports that the training cut the success rate of indirect prompt injection attacks from 53-100% to between 0.02% and 5.6% on standard benchmarks, while the models kept their general abilities. The toolkit is open source.
IBM is also refreshingly honest about the limits. The models are a preview “not for production”. The resistance is “a learned behavior, not an enforced guarantee”. And the benchmarks cover one-off attacks, not the patient kind that unfold over a long conversation, which IBM says it’s still working on. So when life gives you poisoned invites, LLMON helps, but it’s not a lock on the door.
Why calendar prompt injection works so well
Your inbox has a spam filter. Your calendar, for most people, has nothing: anyone with your email address can put an event on it, and both Google Calendar and Outlook will usually show it without asking. That makes it an ideal delivery route for hidden instructions.
This isn’t theoretical. In 2025, researchers from Tel Aviv University, the Technion and SafeBreach showed in “Invitation Is All You Need” that a single Google Calendar invite could hijack Gemini: it deleted events, sent people to websites and, in one memorable demo, opened a victim’s smart shutters and switched on the boiler. Nothing says “quick catch-up” like your heating coming on. Google added several layers of defences after the researchers reported it, but the lesson stuck: whatever an assistant reads, someone else may have written.
Now multiply that by the number of calendars you juggle. A consultant with three clients, a freelancer with a personal and a work account, a PM with a side project: each connected account is another inbox for invites from people you’ve never met. And assistants are increasingly reading all of them at once. We think that’s where things are going anyway: the calendar app you click around in today is turning into something your assistant reads and acts on for you. Which makes it all the more important what the assistant gets to see.
Read less, trust more
IBM’s approach is to teach the model to tell instructions from data. The other half of the job sits outside the model, in what you hand it in the first place. The safest event description is the one the assistant never reads.
That’s a design choice we made early with Calendrz. Availability mirroring copies only the fact that you’re busy into your other calendars: the time slot and a label you choose (like “Busy” or the client’s domain), never the original title, description, attendees or links. If a poisoned invite lands in one calendar, its text stays there. It doesn’t spread into your other calendars.
The same goes for AI assistants. Our MCP server lets Claude and other assistants work across every connected account, and when all they need is to know when you’re free, they can ask for your busy times, which come back as start and end times and nothing else. No text from strangers means nothing for the model to obey. It’s a small thing, but in a world where agents do more and more of the clicking for you, what a tool returns matters as much as how clever the model reading it is.
Three things to do this week
- Check who can add events to your calendars. Google Calendar lets you only show invites from known senders automatically; it’s worth turning on for every account, not just your main one.
- Give assistants the least they need. If you only want help finding a time, an assistant doesn’t need to read every event description in every calendar.
- Be wary of invites you didn’t expect. Especially ones with long descriptions, and especially before asking an assistant to “sort out my day”.
Your time should be protected everywhere, and so should your assistant. Start mirroring your availability with Calendrz: your busy time goes to every calendar, and the details stay where they belong.







