Last month, malicious AI agents compromised 100+ online retailers and stole 600,000+ credit card records. The attackers' total spend: roughly $8,000. Let that sink in. Offense is now automated and nearly free — while most teams still buy defense the old way: expensive consulting, manual reviews, governance as a special project. That asymmetry is the whole game now, and it is getting worse.
The economics are brutal for defenders. The attackers re-run their attack for pennies while you re-run your review board for weeks. Every governance control that requires a meeting, a consultant, or a manual approval is a control that cannot keep pace with an attacker who scripts the whole operation. Until defense is as cheap and automated as offense, defenders lose by design.
The asymmetry is the whole game¶
An $8,000 attack that nets 600,000+ card records is not an anomaly; it is the new unit economics of automated offense. AI agents let attackers scale reconnaissance, exploitation, and exfiltration without scaling headcount. Any defense that scales with headcount — human review queues, ticket-based approvals, periodic manual audits — is structurally outmatched. You cannot hire your way out of an asymmetry this lopsided; you have to automate your way out.
What cheap-by-default looks like¶
Cheap-by-default governance is policy-as-code instead of policy-as-PDF: rules that execute themselves rather than rules that wait for someone to read them. It is spend ceilings that fail closed, so a runaway agent stops itself instead of waiting for a human to notice the bill. It is a per-run tool budget on every agent — a number the runner enforces before the first tool call, not a dashboard someone checks on Fridays. Each of these costs nearly nothing to operate once built, and each one works at machine speed, which is the only speed that matters against automated offense.
The decoration test¶
Apply one test to every control you run: if a control costs more to operate than the attack it stops, it is decoration. Expensive consulting that produces a PDF nobody enforces, manual reviews that take weeks while attacks take hours, governance theater that impresses auditors but cannot stop a loop — all of it fails the test. Keep the controls that operate cheaply and enforce themselves; replace the rest with automated versions or admit they are not controls at all.
The takeaway¶
Governance has to become cheap-by-default — policy-as-code, fail-closed spend ceilings, per-run tool budgets — because offense already is. The teams that win the next phase are not the ones with the biggest security budgets; they are the ones whose controls run at machine speed for machine cost. So here is the question: what is the one agent control you run that actually survives contact with real tool traffic — and which one did you have to rip out?