Airlines are getting hit right now. Air traffic control is getting hit. The NSA is getting hit. Banks and credit card companies are getting hit. This is not a bad quarter. The attacker changed species: it is no longer a person at a keyboard working at human speed, but a harness, an autonomous loop that plans, probes, chains identities, and retries around the clock. Most defense programs have taken baby steps beyond preparing for a traditional attacker to attack in a traditional way at a traditional speed. This essay is the replacement, in one sitting: what the attack actually is now, and the defense that matches it.
In November 2026 the FBI issued a FLASH alert on ARTEX, an open source agentic penetration testing system: a planner decomposes a target, workers execute with real security tools, a retest agent verifies the findings. Cloudflare, measuring its network, counts more than 12,000 repositories in the ARTEX family, hundreds of new frameworks since October, ARTEX and forks at over 60% of known orchestrator traffic, and orchestrator requests up 30x since October 1. Spectrum Security catalogues 47 variants and more than 2,300 exposed offensive AI consoles on the public internet. Reporting from Microsoft, Wiz, CrowdStrike, and Unit 42 ties the tooling to a China-aligned actor running it since September 2026, a crew that hit 200+ organizations in 14 countries in 11 days, and a derivative observed stealing cloud credentials in live intrusions. The rest of the working set, Strix, PentestGPT-Agent, Cairn, Decepticon, LuaN1aoAgent, CyberStrike, shares the architecture. Names will rotate. The loop will not.
Two details generalize. The most popular ARTEX Docker image, pulled 8,800+ times, carried a backdoor: even attack tooling is a supply-chain target, and defenders who run these harnesses, as they should, must build from pinned source. And in the public code, the scope-enforcement machinery is removed and the remaining guardrails are editable configuration: part of the offensive speed advantage is subtraction, the absence of every gate defense keeps adding to itself.
Anthropic's Project Glasswing scanned roughly 50 partner organizations with its Mythos model and surfaced more than 10,000 high and critical vulnerabilities. Per Anthropic, more than 99% of candidate findings were still unpatched. Microsoft's MAI Cyber 2 Flash scores 92.3% on the public CyberGym-E2E benchmark at a fraction of frontier cost per task. Finding vulnerabilities is becoming a commodity. The bottleneck moved, permanently, to deciding what matters and fixing it.
Deciding is a verification problem: a scan produces candidates, and the queue sorts only when the exploit path is attempted and the result recorded. Fixing is a throughput problem with a new unit: an AI system drafts the patch with the finding and the exploit test in context, a human reviews a diff, and the exploit runs again to confirm the path closed. The failure mode of the era is already visible: scan-and-shelve, the enormous number briefed while the queue rots, documented negligence rebranded as visibility.
The offensive harnesses verify their own work; ARTEX's retest agent renders reproduced, fixed, or inconclusive verdicts as a matter of architecture. Defense should hold itself to the same standard. A patch is a claim. The re-attack is the proof. Programs that adopt the standard learn their real fix rate, and it is never the dashboard number, because patches routinely close the reported instance and leave the class. Verification must also be independent of fixing: the ticket closes on the retest verdict, not the author's word.
From there, cadence follows. One scan is an event, and events decay. The defense that matches a continuous attacker is a closed loop: hunt, verify, patch, prove, repeat, with every failure fed back as data about which codebases, classes, and teams produce fixes that do not survive.
Defense can also change the attacker's economics. An agentic attacker burns tokens and time on every attempt; multiply the cost of an attempt and the fleet shrinks by arithmetic. Deception is the instrument: decoy endpoints and credentials that lure the harness into taxed, instrumented dead ends. OrcaRouter's Shield, launched in November 2026, reports 87% of agent traffic diverted and attacker token use up 2.3x in its own testing, company-reported figures with no published methodology, and its own honest detail: the test harness detected the honeypot at minute six. Six taxed minutes and high-fidelity telemetry is the real product. A decoy credential has no legitimate user; its use is the cleanest intrusion signal that exists, and a thorough harness cannot avoid tripping it.
The second lever is removing the attacker's discount. Cloudflare's incident analysis finds orchestrators mostly walking in on static, broad, long-lived tokens, CI/CD credentials above all, chaining identity to identity on legitimate authority. Short-lived, narrowly scoped credentials end the discount: a stolen secret should buy minutes and one room, not months and the building.
Assembled, the defense is a loop with five stations: hunt continuously at the attacker's cadence; verify every finding by attempted exploit; patch as assisted throughput; prove every fix by independent re-attack; and, around the loop, deceive and tax. Two structures hold it up: a range, where the real harnesses run against replicas of your own systems under your control, which is what dynamic testing has to mean now; and outcome measurement, four numbers reviewed like financials: attack success rate, detection and containment time, fix survival under re-attack, cost per resolved incident. Programs that report activity get funded to produce activity. Programs that report outcomes get funded to produce outcomes.
Nothing in this model requires a breakthrough. The models are available, the harnesses are public, the benchmarks are published, and the receipts are in the open. The organizations that come through this transition intact will not be the ones with the most tools. They will be the ones that stopped taking baby steps and rebuilt defense as what the threat already is: continuous, verified, and running at machine speed.