Last updated 16 November 2026
Built for the day someone tries to trick it.
ontopics protects against prompt injection by design, not by hoping the AI spots the trick. It keeps apart the three things an attack needs: your private data, untrusted messages, and the ability to act. Security researcher Simon Willison calls that combination the lethal trifecta.
ontopics never puts all three in one AI
Why AI assistants can be hijacked
Language models can’t reliably tell your instructions apart from instructions hidden inside the content they read. So an assistant that reads your email and can send email can be told, by an email, to send your data somewhere else.
Prompt injection is ranked first in OWASP’s Top 10 risks for LLM applications. Filters that catch most attacks aren’t enough when one success is all an attacker needs. A plain guide to prompt injection →
How ontopics keeps them apart
ontopics makes sure the part that reads untrusted messages can’t act, and the part that acts never reads them. Every action crosses the guard.
Untrusted side · on your Mac
- MessagesWhatsApp, email, websites
- 1 · FilterUnlisted chats and senders dropped
- 2 · ReaderLocal AI. No tools, no internet
- Facts onlyDates, amounts, asks, flags
only facts cross this line, never message text
Trusted side
- 3 · PlannerClaude or ChatGPT. Proposes actions
- 4 · GuardChecks the topic’s list and caps
- 5 · Face IDFor anything that matters
- ActionLogged in Activity (7)
The seven layers
Filter before any AI
Connectors drop everything outside a topic's sources before a model sees it. Unticked WhatsApp chats and unlisted senders never reach the AI at all.
The reader can't act
A local AI on your Mac reads the raw messages. It has no tools and no internet access. Its only output is a fixed set of facts: dates, amounts, what's being asked. Instructions it finds are quoted and flagged, not followed.
The planner never reads the original
Your chosen AI (Claude or ChatGPT) plans from those facts alone, each linked to its source. A hidden instruction never reaches it as an instruction. This follows the dual-LLM approach, refined by Google DeepMind’s CaMeL research.
The guard checks every action
Each action passes through a gateway that checks it against the topic's lists: allowed actions, destinations and spending limits. There is no path from the AI to a tool that skips it.
Face ID for anything that matters
Orders, payments and messages on your behalf need an approval signed by your iPhone's Secure Enclave, behind Face ID. Your Mac checks the signature. A reused approval, or one for a changed amount, is rejected.
The AI never holds your passwords
Logins live in your Mac's Keychain. The AI only ever sees a label like “your Ocado login”; ontopics fills the real one in directly.
A record that can't be quietly edited
Every allowed, blocked and pending action is written to a tamper-evident log on your Mac, shown in the Activity screen.
See it stop an attack
Walk through a real injection attempt, step by step.
Try the prompt-injection demo →
What ontopics can’t protect against
We’d rather say this up front.
- A compromised MacIf someone already controls your Mac, they control what runs on it.
- Approving something you shouldn'tFace ID proves it was you, not that it was a good idea. Approval screens show exactly what will happen so you can judge.
- A reader that misses somethingThe local AI can miss or misread a fact. Every line links to its source so you can check.
- Changes we don't controlWhatsApp's local format isn't documented. When it changes, ontopics stops and tells you rather than guessing.
Closed source, open design
ontopics isn’t open source during the beta. Instead, we publish this design in full, show every action in Activity, and plan an external security review before a wider launch.