Posted in Security Operations
The prompt is not the boundary. Execution is. A model can be talked into almost anything. Better…
Schomer — Principal Security Solutions Architect & Founder
The prompt is not the boundary. Execution is.
A model can be talked into almost anything. Better instructions, a cleverer system prompt,
another warning in bold - none of that changes what the model is. It produces proposals. Text
in, text out.
So stop asking the prompt to be the security control.
Put the control where the action happens: at execution. Every proposed action gets checked
before it runs, and the gate - not the model - decides:
Allow - the action matches the rules, it runs.
Deny - it breaks a rule or matches no rule, it does not run.
Escalate - it is state-changing or ambiguous, a human decides first.
That is how I run it in my own lab. On Modbus writes, a function code 16 write carrying values
[52, 53] was blocked by the gate. The same write carrying [52, 52] was forwarded. Same
model, same prompt, same request shape - different values, different decision, made at the
gate.
On shell commands the same pattern held: a read was allowed, a destructive command was
denied at authorize, and a privileged command with no rule for it was denied on default-deny.
The model did not get safer. The boundary moved to the only place it can actually hold - the
moment before something executes.
If your agent can act, ask one question: who decides at execution - the model that was just
manipulated, or a gate that cannot be talked out of its rules?
- Justin Schomer, Eternal Boundaries
3 reactions · 0 comments