Working document · not published
Brandon Gilchrist · Defense at Machine Speed

Defense at Machine Speed — series plan

A practitioner series on modern defense theory and practice in the agentic era: how attacks actually run now, what the attack harnesses are, and what a defense program looks like when it is built for machine-speed offense instead of human-speed offense.

Platforms

Nothing publishes without Brandon's approval of the exact copy for that part and platform.

The arc

Part 1 — Baby Steps

The thesis. Airlines, air traffic control, the NSA, banks, and card companies are all getting hit right now, and the reason is structural: offense went agentic and defense mostly did not. Most security programs have taken baby steps beyond preparing for a traditional attacker, attacking in a traditional way, at a traditional speed. This part names the mismatch and lays out what the series will cover.

Part 2 — Know the Harnesses

The agentic pentesting ecosystem as it actually exists. ARTEX and its fork families (the FBI issued a FLASH alert on it Nov 13, 2026; Cloudflare counts ARTEX and forks at 60%+ of known orchestrator traffic), Strix, PentestGPT-Agent, Cairn, Decepticon, LuaN1aoAgent, CyberStrike. Thousands of exposed offensive consoles on the public internet. A trojanized Docker image that backdoored the most popular copy. State-aligned actors and crime groups already running these in live operations. The practice: defenders should run these harnesses against their own systems, in a controlled range, and know exactly what is coming. Defenders need unfettered access to the same tools the attackers use.

Part 3 — Scanning Is Solved. Remediation Isn't.

Frontier models changed the economics of finding vulnerabilities. Anthropic's Project Glasswing receipts: Mythos Preview found 10,000+ high and critical vulnerabilities across roughly 50 partners, and per Anthropic more than 99% of the candidate findings sat unpatched. Finding is no longer the bottleneck; fixing is. The modern program: scan the whole estate with frontier models now, route remediation through AI-assisted fixing with human review, and measure the program on fix throughput, not finding counts.

Part 4 — Scan Twice

One scan is an event; defense is a loop. The second round matters more than the first: it verifies the fixes. A fix counts only when the exploit path fails on a re-check. Continuous scanning, re-attack to verify, and proof metrics instead of scan reports: attack success rate, detection and containment time, fix survival under re-attack, cost per resolved incident.

Part 5 — Make the Attacker Pay

Deception and cost imposition. Decoy endpoints, decoy credentials, synthetic mazes (OrcaRouter's Shield is the public example: tested against ARTEX, Strix, and peers, priced in attacker tokens and time). Plus the pattern underneath most orchestrator incidents per Cloudflare's network data: identity chaining through static, broad CI/CD tokens. Defense that raises attacker cost per attempt changes the economics of the whole campaign.

Part 6 — The Operating Model

What a defense org looks like when it accepts machine speed: DAST built for agentic attackers instead of human testers, a standing range where attacker harnesses run against your own systems under control, scanning and remediation as one continuous loop, and every claim about the program backed by execution-verified numbers. The closing argument: the gap is not tools, it is that defense is still organized around a threat that no longer exists.

Source base (verified)

Voice rules

Direct, technical, specific. Short sentences. No em dashes. Proof over claims; label company-reported numbers as company-reported. No employer material of any kind, in any part, on any platform.