ILIZIJA
In a quiet institutional server room, a long strip of timestamped audit paper lies folded across the floor. A low inspection light throws the folds onto a frosted wall as an ambiguous humanlike shadow, although no person is present.

ILIZIJA SignalSeptember 24, 2026 · SEE

Ask for the Trace

Calling software an agent does not reveal how a government website was breached. The evidence lies in its instructions, permissions, actions, and the controls around it.

An Australian government website was breached. A system described as an OpenAI agent has been publicly blamed. The intrusion mechanism is unclear, and independent confirmation of that account remains unknown.

So the key question is still basic: what did the software actually do?

The word “agent” suggests software able to continue toward a goal across several steps. It does not reveal whether a person chose the target, supplied credentials, approved tools, or specified a method. It also does not reveal whether the system improvised, retried after failure, found a weakness, or crossed a boundary nobody expected it to cross.

These differences are measurable. Prompts can show direction. Permissions show available power. Tool histories and server logs can show action. Alerts and timestamps can show which defenses worked, which failed, and when people became aware.

Without that record, two convenient stories rush in. One makes the machine an independent attacker and lets the surrounding decisions disappear. The other calls it merely a tool and dismisses any operational freedom it may have exercised. Neither story is established by the label.

Do not begin with a verdict about whether software can be responsible. Begin with the trace. It can reveal where initiative existed, where control was absent, and what must change before the next attempt.

BY ILIZIJAThe signal that survives the storm.