Agentic AI Just Moved Into Your Living Room
A viral threat model for AI-driven smart homes shows why letting LLMs act on your behalf inside physical space needs new security thinking.
The signal: A widely-shared threat model breaks down how AI-driven smart home devices — voice hubs, cameras, agentic assistants — create attack surfaces that didn’t exist when these things just followed hardcoded rules.
Why it matters: Once an LLM is deciding what your thermostat, lock, or camera does next, you’ve shipped a decision-making system into someone’s physical space, not just a feature. Prompt injection, misread context, or a compromised skill now has real-world consequences — heat, light, locks, cameras.
The pattern I’m watching: Agentic AI is leaving the browser and the API sandbox and walking into physical environments — this same week’s autonomous umbrella is the cute version, smart home is the consequential one. Security models built for deterministic firmware weren’t designed for probabilistic decision-makers, and most teams are shipping the AI layer faster than they’re threat-modeling it.
What I’d do: If you’re building anything agentic that touches hardware, write the threat model before you write the skill — enumerate what happens when the model is wrong, tricked, or ambiguous, not just when it’s right. Add explicit human confirmation for any action with physical or financial consequence, and log every autonomous decision like you’ll need to explain it to a lawyer someday.