Quick Answer
Attackers now have AI tools that compress weeks of reconnaissance into hours, and in several documented cases reached privileged cloud access in under three hours. We believe attacks will keep accelerating through 2027 as this technology matures, which is why we are researching these offensive AI frameworks directly: to understand what is already in the hands of attackers, to prioritize where we invest in our own services next, and to give our clients continuous penetration testing that keeps pace with how fast the threat landscape is moving.
A few weeks ago Microsoft published something that brought this home for us. In several real intrusions this year, attackers went from first touching a network to having privileged access in the cloud in under three hours. Not three weeks. Three hours. A few days later the FBI issued an advisory describing the same underlying tooling, independently. We have been in this industry long enough to be skeptical of “AI changes everything” claims. This one holds up, and it is the clearest argument we have seen yet for why continuous testing has stopped being a nice-to-have and started being the baseline.
What’s Changed
There is a small family of open-source tools now, including ARTEX, Cairn, and CyberStrikeAI, that do something genuinely new. Give one a goal, something like “get into this network,” and it plans the attack itself, runs it, reads what comes back, and decides what to try next. No human deciding the next step. Several ship with the “ask a human before doing anything risky” setting switched off by default. Every one of them started life as a legitimate security research tool, and the uncomfortable part is that the same capability built to help a tester find a weakness first is now running the attack itself, continuously, for anyone who downloads it.
The Proof Is Already In
This is not theoretical. Microsoft attributes a real ConnectWise supply chain compromise to a group using one of these tools, reaching over 2,250 customer tenants through a single trusted vendor. Researchers separately caught another actor working through more than 100 companies methodically, rather than picking obvious targets. Check Point documented a single operator, not a nation-state team, just one person, running over 5,000 AI-executed commands against government systems and walking away with nearly 200 million records. Different tools, same pattern: attackers are not smarter than they used to be, they are just not the bottleneck anymore. The AI is.
Why A Once-A-Year Test Can’t Keep Up
Official advisories on this activity say it plainly: these tools go after partner portals, legacy systems, and third-party integrations, not the flagship application everyone has already hardened. The ConnectWise breach was not a front door being kicked in. It was a trusted vendor’s door, left unwatched.
That is exactly where a once-a-year penetration test is weakest. A scheduled test has a scope and an end date. An AI agent does not. It can work through every overlooked asset in an organization’s attack surface, continuously, at a pace no human team is paid to match, and it will find the one thing everyone forgot about faster than the next scheduled test would have.
Why We’re Researching This Space
Here is where we will be honest about where the industry stands: attackers currently have better AI tooling than most defenders do. That is not a comfortable thing to admit, but pretending otherwise does not help anyone. The tools above prove the offensive side has figured out how to put AI to work at scale. The defensive side is further behind than it should be, and the gap is widening quickly. A lot of companies, our clients included, stand to benefit from AI-driven testing as part of what we offer, and closing that gap for them is exactly why we are investing here.
That is why we are putting real effort into researching this space properly, not to write about it from a distance, but to understand how these frameworks actually work. For a continuous testing company, that research is not optional. If we do not know what is already in the hands of the people attacking our clients, we cannot make good decisions about where to invest next. It is how we prioritize our own roadmap: which capabilities matter now, which are still a year away, and where the real gap sits between what attackers already have and what defenders are using. Things are moving at a rapid pace in this space, and we intend to keep up with it rather than react to it after the fact.
That same research is also what lets us build equivalent capability into how we test and protect our clients, rather than guessing at it. If an AI agent can run continuous, autonomous reconnaissance against a target, there is no good reason an organization should not have an equivalent working for it instead of against it.
What We’d Ask You To Take From This
We are not writing this to create alarm, and we are not going to pretend the answer is “buy more tools.” The honest message is simpler: the gap between how often attackers now test an organization and how often it tests itself has widened, and it widened fast enough this year that last year’s cadence is already behind. If you are not sure whether the overlooked parts of your estate would survive something that tests them continuously rather than once a year, that is the conversation worth having before 2027, not after.
Frequently Asked Questions
What are agentic AI attack frameworks like ARTEX and Cairn?
They are open-source tools that pair a large language model with hacking workflows, so an attacker can give the AI a goal and it will plan, run, and adapt the attack itself with little human input.
How fast can an AI-driven attack reach privileged access?
Microsoft and the FBI have both documented cases where attackers reached privileged cloud access in under three hours from first targeting a network.
Why can't annual penetration testing keep up with AI-driven attacks?
A scheduled test has a fixed scope and end date, while an AI-driven attack can work through every overlooked asset continuously. The gap between test cycles is exactly where these tools are designed to operate.
Is Rootshell researching AI-driven attack tools?
Yes. We are researching frameworks like ARTEX, Cairn, and CyberStrikeAI directly, so we understand what attackers already have and can prioritize our own service roadmap and client-facing capability accordingly.
What should a board be asking about AI-driven attacks?
Whether the overlooked parts of the estate, such as partner portals, legacy systems, and third-party integrations, would survive continuous testing rather than an annual one, and whether the current testing cadence still matches how fast attackers are moving.
Is continuous penetration testing enough on its own?
Continuous testing closes the gap that point-in-time testing leaves open, but it works alongside strong exposure management and fast patch and detection cycles, not as a replacement for them.

