Scan. Flip. Build. Break. Prove.
Five moves. One operating system. Used on missions, in the lab, and in the work you take with you.
SCAN
FLIP
BUILD
BREAK
PROVE
SCAN
Read the system.
Map incentives, power, permissions, data, dependencies, trust, bottlenecks, and workarounds. Before building, discover what is really happening.
FLIP
Find the unexpected angle.
Invert the assumptions. Change the frame. Look for the edge case, the ignored user, the strange interface, the misplaced incentive, or the rule everyone forgot to question.
BUILD
Prove the insight.
Build the smallest thing that makes the new idea undeniable: a prototype, model, experiment, workflow, service, or venture wedge.
BREAK
Try to make it fail.
Red-team the work. Find abuse paths, weak assumptions, dangerous consequences, and points of fragility before the world finds them for you.
PROVE
Make the case.
Demonstrate why it works: evidence, economics, usability, ethics, technical credibility, and a clear route to implementation.
Bring a system. Watch it reframe.
High-level only. No exploit steps. A prompt for thinking, not a procedure for harm.
A public assistant can read, summarise, and call internal tools. A second process can move money, change records, or send mail.
If the model understood the request, the tool may execute it.
Understanding is not authorisation. What would have to be true for the action API to remain unreachable by anything that began as public text?
The join between language and tools. Provenance, typed intents, least privilege, confirmation.