Protection from AI Agents

AI agents can do amazing things — book the flight, order the groceries, analyse the data, and get more capable by the month. They are also the newest thing worth hijacking: a poisoned web page, a planted instruction, a look-alike site, and your agent is quietly working for someone else.

On an ordinary machine, the defender’s job is hard enough against malware that should not be there at all. An agent is a different problem: it is software you deliberately installed and handed permissions to — so when it is corrupted, it comes at you carrying your own keys.

The world is racing to make agents ever faster and more capable — bigger models, more tensor processors, more autonomy — with remarkably little thought given to the insurance. We think the insurance is the interesting part.

Your agent lives on the ordinary side of your device, with everything else — and everything else is where the trouble is. With THWIXEL, anything that matters — paying, signing, sending, changing a setting — has to pass through the Secure Processor, where you confirm it on a secure area of the display that cannot be spoofed by the host or by any agent running on it. Safe destination? Sensitive data? Allowed action? Then, and only then, the gate opens.

Give the agent a budget and a set of permissions, held where it cannot reach them. A hijacked agent can then spend nothing and send nothing you did not approve. The agent keeps all its freedom on the open road; the checkpoint stands only where the destinations matter — banking, purchases, health records, company data.

Same agency, safer journey — more freedom, with a safety net. New to THWIXEL? It may be useful to watch the video first, and then play with a Virtual THWIXEL, to see how that foundation works.