Agent-like Wigner's friends
Can you get information into or out of a quantum box without leaving a mark?
Map / survey
In the Wigner's friend story, a "friend" inside a sealed lab measures a quantum system, and an outside observer later treats the whole lab as one quantum object. Laux and Cavalcanti (arXiv 2609.12527) make the friend agent-like and run the setup on IBM quantum hardware.
The lesson we took from it: every record leaves a mark. Moving information in or out of the box disturbs it. Undoing the friend's measurement, a rewind, only works if nothing leaked. So a rewind that passes shows the box stayed sealed.
Record, leak, rewind, check
θ = 60°
Each trial: the system starts in a 50/50 superposition. The friend records it into memory. A stray leak nudges an environment qubit by angle θ, but only when the system is 1. The friend's record is then rewound, and a check asks: is the system back in its starting state? It passes with chance (1 + cos(θ/2)) / 2. At θ = 0 every rewind passes.
In plain words
The observer effect is usually told as a nuisance. Here it is a feature. Because any copy that escapes the box changes what a rewind can do, a clean rewind certifies that no copy escaped. That is the lock.
It also points the other way. Harvest-now-decrypt-later attackers store encrypted traffic today and wait for a quantum computer to read it. And quantum key distribution (sharing a secret key using single photons) has been beaten by detector blinding, where bright light forces the receiver's detectors to click on command. We called that a kind of quantum prompt injection: the attacker steers the measurement instead of breaking the math.
Prior artThe paper itself.
Next clickProvable forgetting, which became the no-deleting thread.