For years, many organizations have operated under a dangerous misconception: "Who would want to hack us? We are too small to be a target."
This mindset is a relic of an era where hackers were humans looking for specific, high-value prizes. Today, that has changed. Modern cyberattacks are increasingly automated and opportunistic. Attackers do not need a motive to target you; they only need a script that identifies a vulnerability. If your network is reachable, you are a target by default.
The "smallness" of an organization no longer provides a shield; it only provides a lack of preparation.
The Fragility of Interconnection: A Week of Chaos
This week alone has demonstrated how fragile our interconnected systems have become. Whether through malicious intent or simple technical error, the disruption was systemic and immediate:
-
NSA NSANet Outage
The highly classified network the NSA uses for internal communications went down after a failed software patch. -
Microsoft 365 Global Outage
A giant Outlook outage showed the risk of centralized cloud dependencies. -
United Airlines System Failure
A system outage left pilots without final aircraft weight data and delayed departures nationwide. -
Air Traffic Control Disruptions
A nationwide ground stop traced to one malfunctioning router during an FAA upgrade. The FAA has since paused the rollout. -
Banking Connectivity Issues
Online outages hit Chase, Bank of America, and other major banks. Chase blamed a technical bug, not a cyber issue. -
Apple iCloud Outage
An internal update knocked out iCloud sign-in, Apple Pay, the App Store, and even Apple's own employee sign-in and VPN. Apple said it was not a cyberattack.
To be clear, none of these was an attack. The Director of National Intelligence, Jay Clayton, said publicly that the run of outages was not a cyberattack and not a coordinated nation-state effort, and the companies involved blamed their own updates and bugs. That is the point. If routine internal mistakes can take down sign-in, payments, and air traffic in a single week, imagine what a deliberate, automated adversary could do with the same dependencies.
The deliberate version is what the FBI just warned about. It issued an alert on ARTEX, an AI-driven framework that lets attackers move from initial targeting to privileged access in under three hours in some intrusions.
Moving from Static Defense to Dynamic Readiness
If the window for response is shrinking, your defensive posture must expand. We must move away from "compliance-based security" the practice of checking boxes once a year and toward continuous adversarial readiness.
To minimize damage in this new era, organizations should implement three core strategies:
-
Continuous Adversarial Testing (Red Teaming & BAS)
Annual penetration tests are snapshots of the past. You need continuous Breach and Attack Simulation (BAS) and regular Red Teaming exercises. You must act as your own worst enemy to find the holes before an AI does. -
Purple Teaming Culture
Security cannot exist in a vacuum. Defenders (the Blue Team) and attackers (the Red Team) must work in a tight, collaborative loop. The goal is not just to "catch" the attacker, but to use every simulated attack to immediately tune detection logic and automated responses. -
Resilience over Robustness
Since a breach is an inevitability, focus on limiting the "blast radius." This means aggressive micro-segmentation, immutable backups, and automated identity revocation. You cannot prevent every entry, but you can prevent a three-hour breach from becoming a three-month catastrophe.
So how do you put all this in motion? Let me take you through that.
The Logistics of Hacking
Preparing for this new era requires more than just theoretical planning; it requires practical, hands-on simulation. This means using the very tools attackers use to find your vulnerabilities. AI is the QA tester you wish you had, but secretly hope you never have.
However, you will likely encounter a significant hurdle: the ethical safeguards and "safety rails" built into mainstream models like Claude or ChatGPT. When you attempt to simulate an attack or probe for specific system weaknesses, these models are designed to refuse.
AI is the QA tester you wish you had, but secretly hope you never have
To conduct effective, high-fidelity drills, you need models and providers suited to authorized security testing. If a mainstream model refuses to assist in a legitimate, scoped red-teaming exercise, shift your workflow toward more permissive options:
-
Local LLMs
Running open-source models (like Llama or Mistral) on your own hardware removes all external censorship and allows for unrestricted testing of exploit logic. -
Unfiltered API Access
Utilizing providers like OpenRouter to access models with fewer restrictive layers, or specifically targeting models like DeepSeek that may offer a different threshold for technical/adversarial queries. -
Agentic Probing
Don't just ask a single question. Build specialized prompting agents designed to iteratively probe your architecture, simulating the multi-step logic an AI-driven attacker would use to chain vulnerabilities together.
Set the ground rules before you run anything
Permissive tools raise the stakes, so the logistics matter as much as the tooling.
-
Written authorization and scope
Get sign-off from leadership and the owners of every system in scope. Define what is in bounds, what is off limits (production databases, third-party SaaS, partner networks), and who can call a stop. -
Contain the tooling
Run the models and attack tooling on a segmented host or VM, not on a laptop that sits on the corporate network. A local model with agentic access to your network is itself a risk if it is not contained. -
Start small, not perfect
Most organizations cannot afford a full mirror of production, and waiting for one means never starting. Pick one slice that matters, such as the identity provider, a single business application, or one cloud account. Test a cloned or dev instance of it, or test live with tight scope, off-hours timing, and a rollback plan. A narrow test of something real beats a broad test of a toy. -
Safe handling of data
Use synthetic or scrubbed data. Do not send real credentials, customer records, architecture diagrams, or source code to third-party APIs such as OpenRouter. Anything sensitive stays on a local model or a contracted endpoint with no data retention. Treat prompts and outputs as sensitive artifacts. -
Humans in the loop
Agents propose, people approve. Require a human to review any step that changes state, and set hard limits on the actions, rate, and duration of an agent run. -
Full logging
Record every prompt, command, and result. You need the trail to reproduce findings, tune detections, and show auditors what you did.
Split the attackers into two teams
Do not run one red team. Run two, with opposite starting points, because real attackers come in both flavors.
Team One: the insiders.
These are the people who know how the systems work: the engineers, admins, and architects who built them. They know where the shortcuts live, which service accounts are over-privileged, and which "temporary" exceptions became permanent. Give them the permissive AI tooling and point them at those weak spots. They find the deep, specific flaws that an outsider would take days to discover.
Let this team point the AI at the source code, using a local model or a no-retention endpoint, never a public API. Have it review the code for flaws, then try to exploit what it spots against a dev or tightly scoped target. This confirms whether a hole is real. It also shows whether something else is hiding it. Often the vulnerability is genuinely there, but a ring fence such as Cloudflare, a WAF, or a rate limiter is stopping the exploit. That is a useful result, but it is a fragile one. The flaw is still in the code, and your protection now depends on one layer you do not control. Log each of these as a hole that is masked, not a hole that is fixed, and schedule the code fix anyway.
Team Two: the cold prompters.
These people do not know how the systems work, and that is the point. Give them the same AI tooling and the same scope, and have them prompt-hack from zero: what is exposed, what can I enumerate, what can I chain. This mimics ARTEX-style automation, which arrives with no inside knowledge and no patience for your org chart. It also tests something the insiders cannot: what the AI finds on its own, without a human steering it toward the answer.
Keep both teams off your own network.
Run the attacks from outside, such as a cloud VM or a separate connection, not from a laptop on the corporate LAN or VPN. An attacker does not start with a trusted internal address, and a test from inside skips your perimeter, your external monitoring, and your rate limits. If the only way in is from the inside, that is a finding, not a shortcut. Brief a small white cell in advance: the few people who can call a stop, and your provider if needed. Give them the source addresses and the schedule. Keep your SOC blind, or your time-to-detect numbers will be fiction.
The comparison is the payoff:
-
Found by both
Your obvious exposure. Fix these first, because an automated attacker will find them within minutes. -
Found only by insiders
Weaknesses that depend on knowledge. Do not treat these as low priority. AI closes the knowledge gap fast: an attacker who steals a repo or lands a foothold gets insider knowledge within minutes. They rank below the shared findings only because they need that first step. -
Found only by the cold team
Blind spots. Your own people were too close to see them, and this list is often the most humbling.
Rotate people between teams from one round to the next, so that today's insider becomes tomorrow's cold prompter on a different system. Small shops can do this with two people, or with one person and an agent running cold.
Make the drill measurable
A simulation only helps if it produces numbers you can improve. Track:
-
Time to detect
How long from first action to the first alert? -
Time to contain
How long until the affected identity or host is isolated? -
Coverage
Which steps in the chain did your tooling see, and which went silent? -
Fix cycle
How many days from finding to remediation, and was it retested?
Against a three-hour attacker, these are the figures that matter. If your detection time alone is longer than the window, you have found your first priority.
Build a cadence
Run a small, focused exercise every month and a larger multi-step chain every quarter. Re-run after any major change to identity, network layout, or cloud configuration. Feed each result back to the Blue Team the same week, which is the purple team loop in practice.
If you aren't willing to break your own defenses, you won't be ready when an adversary does it for you.
The Mandate for Leadership
The era of "security as an afterthought" is over. In an environment where attackers operate on a sub-three-hour timeline, security is no longer a technical requirement; it is the fundamental prerequisite for operational continuity.
Stop asking if you will be breached. Start asking how fast you can recover when you are.
If you are looking for some assistance for getting started then talk to us at ThinkLikeaACTO.ai
PS AI Disclaimer Gemini recreated the top elite hacking squad from The LadyKillers (2014), with me as part of their team



