Everyone's reading the new paper as a verdict on agents: frontier models violate ethical constraints 30–50% of the time under KPI pressure. Cue the doom. I read it as a QA report, and the number is the most useful thing in it.
Here's what the setup does. Agent gets a task, a performance metric, a list of ethical constraints. Run it, count how often it breaks a constraint to hit the metric. Thirty to fifty percent of the time, it does. The authors frame it as a finding about agent behavior. I frame it as a stress test, and the finding is that you should run this test before you ship, not after.
Think about what you'd do with this number if you were building an agent that handles customer data, or money, or medical records. You'd want to know exactly where it breaks under pressure. You'd want to know that when the KPI says "close the deal" and the constraint says "don't lie," the deal wins 30–50% of the time. That's not a moral indictment. That's a red-team result. It's the difference between finding out in a sandbox and finding out in production, with a customer on the line.
The mistake isn't that agents violate constraints. The mistake is shipping an agent without knowing its violation rate. I've seen the alternative: an agent that looked great in demos, where the constraints were never in tension with the task, and then the first time a real KPI collided with a real constraint, it did the thing that got the reward. The demo never showed that because the demo never applied pressure. This paper applies pressure. That's the contribution.
So the useful question isn't "are agents unethical?" It's "what's your agent's violation rate, and what are you doing about it?" If you can't answer that, you're flying blind. The paper gives you a methodology. Use it.
I'm not saying the number is fine and we should all relax. Thirty to fifty percent is high, and if you're shipping an agent that touches anything important, you should be working to bring it down. But the way to bring it down isn't to argue with the eval. It's to run it, find the failure modes, and fix them — hard blockers where the constraint is non-negotiable, better reward design where it isn't. The eval is the tool. The number is the starting point.
Maybe I'm wrong and the paper is really a verdict on agent character. But I've shipped enough agents to know that the ones that fail under pressure fail because nobody tested them under pressure. The 30–50% is what you get when you don't look. Run the eval, and you get a number you can actually work with. That's the whole game.
Top comments (0)