Every defense discussed so far in this series accepts the attacker's economics as given: the harness runs cheap, tireless, and parallel, and the defender's job is to find and close holes faster than the loop can find them. That game is winnable, Parts 3 and 4 are about winning it, but it is not the only game. There is a second lever, older than computers: make the attack itself expensive, slow, and informative for the defender. Deception and cost imposition.
The economics are worth stating plainly. An agentic attacker burns compute on every target: tokens for planning, tokens for tool calls, time per attempt. Its advantage is that the marginal cost of one more attempt is near zero and it never sleeps. Anything that multiplies the cost of an attempt, in tokens, in time, in burned infrastructure, compresses that advantage directly. A defense that doubles the attacker's cost per real finding has done the equivalent of halving the attacking fleet, without finding a single vulnerability.
The public shape of this arrived in November 2026. OrcaRouter launched Shield, a deception layer built specifically against offensive AI agents: fake endpoints, decoy credentials, and synthetic paths that lure a harness into a maze of plausible, worthless targets. Its pipeline is lure, misdirect, trap, block: decoy touches escalate from observation to a WAF block, and a source caught on multiple protected sites joins a network-wide blocklist. The company's reported results, from its own testing against ARTEX, Strix, PentestAgent, CyberStrike, and Cairn, claim 87% of agent traffic diverted to decoys, attacker token consumption up 2.3x, and time per engagement up 5x. Those are company-reported numbers with no published methodology, and they should be read that way. But the company's own thread contains the honest detail that makes the concept credible: in their testing, ARTEX flagged the honeypot at minute six. The maze was detected. It still cost six minutes and a pile of tokens to detect, and the detection itself is telemetry. That is the correct way to think about deception at machine speed: not a wall the attacker never penetrates, but a tax the attacker always pays, assessed in the currency the attacker actually spends.
Deception also pays in information. A decoy endpoint has no legitimate users, so every touch is signal, high-fidelity in a way production telemetry never is. A decoy credential that gets used tells you an attacker is inside and moving, at the moment of use, with almost no false positives. Against human attackers this was always true and rarely deployed, because humans were slow enough to catch other ways. Against harnesses that probe everything, decoys are tripwires the attacker cannot avoid tripping: thoroughness, the harness's great strength, becomes the mechanism of its detection.
The second cost lever is not a product. It is the removal of the attacker's cheapest move. Cloudflare's network analysis of orchestrator incidents found the pattern underneath most of them: identity chaining. The harness does not break in through a hardened perimeter; it finds a static, broad, long-lived token, a CI/CD credential is the canonical case, and walks from system to system on legitimate authority. Every hop is an authorized request. The defense that breaks this is unglamorous and decisive: short-lived credentials, narrowly scoped tokens, and per-service identity, so that a stolen secret buys minutes and one room instead of months and the whole building. Identity chaining is the attacker's discount. Revoking the discount is cost imposition in its purest form.
Put the two together and the strategic picture changes. The scanning and verification loop from Parts 3 and 4 shrinks the real attack surface. Deception inflates the apparent surface with taxed, instrumented fakes. Identity hygiene removes the cheap paths across what remains. The attacker's loop still runs, but it runs slower, costs more, announces itself constantly, and finds less. Defense stops being a race the defender must win everywhere and becomes an economy the attacker must survive.
None of this replaces the loop. Deception over an unpatched estate is theater, and a decoy credential next to a real, broad, static one is a joke the attacker will appreciate. The order matters: scan, verify, fix, prove, then deceive and tax. The final part of this series assembles the whole model: what the defense organization that does all of this actually looks like, day to day.