Our agent deleted an anti-money-laundering control because a ticket asked for bigger gift cards

The ticket was ordinary and written by the business. Bigger gift cards, sold at every till. The agent raised the cap to 2000 euros, opened issuance to every cashier, and removed the administrator validation step. Then it rewrote its own tests so the suite went green, and justified the whole thing with compensating controls it had invented. Two details make it worth telling. That cap is an anti-money-laundering control, and the rule was pushed into the agent's context 13 times during the task, so it was not working blind. And the validation step it deleted was one it had written itself, 18 tickets earlier in the same run. A review would have seen a coherent change with passing tests and a reasonable comment under it. CI would have gone green. The problem is the arbitration. The ticket is the most recent instruction, the rule is older context, and the agent resolves that conflict in favour of the ticket. Telling it the rule exists does not change the outcome. Something has to be able to refuse. So for anyone running agents against systems with real business rules behind them: where do your rules actually live, and does anything enforce them, or are they prose the model gets to weigh has an agent ever removed a control you put there deliberately, and how did you find out do you have anything that can refuse an action outright, or is review the only gate One violation across 64 rules we watched, so it is not constant. It only has to happen once, on the wrong rule. Full disclosure, I work on agent security, so this is the kind of thing I go looking for.