Part 5 of Defense at Machine Speed: Make the Attacker Pay.
Parts 3 and 4 accepted the attacker's economics: the harness runs cheap, tireless, and parallel, and defense wins by finding and proving fixes faster. There is a second lever. Make the attack itself expensive, slow, and informative.
An agentic attacker burns compute on every target: tokens for planning and tool calls, time per attempt. Anything that multiplies the cost of an attempt compresses the attacker's core advantage directly. Double the cost per real finding and you have halved the attacking fleet without patching anything.
The public shape arrived this month: OrcaRouter's Shield, a deception layer built against offensive AI agents, with fake endpoints, decoy credentials, and synthetic paths that escalate from observation to blocking. The company reports 87% of agent traffic diverted and attacker token use up 2.3x in its own testing against ARTEX, Strix, and peers; company-reported, no published methodology, read it that way. The credible detail is their own admission that ARTEX detected the honeypot at minute six. Detection still cost six minutes and a pile of tokens, and the attempt itself is telemetry. Deception at machine speed is not a wall. It is a tax, assessed in the currency the attacker spends, and decoys double as tripwires: a decoy credential has no legitimate user, so its use is a high-fidelity alarm, and a thorough harness cannot avoid tripping it.
The second lever is removing the attacker's cheapest move. Cloudflare's analysis of orchestrator incidents found the pattern underneath most of them: identity chaining on static, broad, long-lived tokens, CI/CD credentials above all. Short-lived, narrowly scoped credentials revoke that discount. A stolen secret should buy minutes and one room, not months and the whole building.
Order matters: scan, verify, fix, prove, then deceive and tax. Deception over an unpatched estate is theater. Full piece on Substack.