Ben Bernstein, Manager of Cybersecurity Advisors at Huntress, is a cybersecurity professional with more than a decade of experience spanning technical support, systems administration, client success, technical account management, and security advisory leadership. He began his career in hands-on IT roles, progressing from helpdesk work to systems administration before moving into account management and client success at Integris. Bernstein later spent three years at Red Canary, where he managed relationships with enterprise customers as a Technical Account Manager and Senior Enterprise Technical Account Manager. Since joining Huntress in 2024 as a founding Technical Account Manager, he has advanced through technical account management leadership and now manages the company’s Cybersecurity Advisors, drawing on his combination of technical expertise and customer-facing experience to help organizations better understand and respond to evolving cyber threats.
Huntress is a cybersecurity company providing a fully managed security platform designed to protect organizations across endpoints, identities, logs, and users. Its platform combines technologies including Managed Endpoint Detection and Response (EDR), Identity Threat Detection and Response (ITDR), Security Information and Event Management (SIEM), and security awareness training with a 24/7 AI-centric Security Operations Center (SOC) staffed by human security experts. Huntress says its technology currently protects more than 5 million endpoints and 15 million identities across more than 277,000 businesses, with its team handling threat detection, investigation, response, and remediation on behalf of customers.
You started your career in helpdesk and systems administration before moving through technical account management and ultimately into cybersecurity advisory leadership at Huntress. How has that hands-on IT background shaped the way you think about security today, particularly when the same tools administrators rely on can also become an attacker’s entry point?
I am eternally grateful for my IT background every single day now that I am on the dedicated cyber side of the house. It gives me a practical perspective. When you know how operating systems, applications, and system processes behave when everything is working normally, the anomalies and threats really stand out. My time in IT also built a troubleshooting and root-cause mindset. You learn quickly that you have to follow the thread to make sure you are fully scoping an issue or a threat.
I remember relying heavily on the same tools, like RMMs, that get abused by threat actors today. Having deployed them myself, I fully appreciate their power in both an IT admin’s and an attacker’s hands. Because RMMs are everywhere, it puts pressure on security teams to look at all the activity surrounding their execution. Having an RMM run on a machine is not inherently evil. But if that RMM suddenly starts dropping additional tooling or spawning unfamiliar scripts, it forces defenders to stop relying on trusted software lists and focus entirely on observing behaviors.
AI is often discussed in cybersecurity in terms of sophisticated autonomous attackers, but how much of AI’s immediate impact is simply making ordinary cybercriminals faster and more capable? Are we reaching a point where relatively low-skilled attackers can execute campaigns that previously required much deeper technical expertise?
There is a lot of panic right now about autonomous AI hacking systems, but the reality on the ground is that this is not novel tradecraft. These models are not breaking into networks using highly sophisticated techniques. They are doing exactly what human script kiddies have done for years by scanning for internet-exposed, unpatched, and misconfigured assets. It makes headlines because the phrase “autonomous AI” sounds scary, or sexy depending on who you ask, but the actual mechanics of these attacks are basic.
The actual problem we face is that AI lowers the barrier to entry for cybercriminals. Attackers can automate the tedious parts of an exploit and scale their operations without needing deep technical expertise. It allows low-skilled attackers to execute campaigns much faster, which puts a lot of pressure on businesses. You can no longer install a security tool and forget about it. Organizations have to assume their networks are constantly being probed.
Huntress has observed attackers using LLM-generated infostealer scripts during RMM intrusions. What does this tell us about how generative AI is changing the economics and accessibility of cybercrime?
It tells us that generative AI is streamlining the operational overhead of cybercrime. Writing an infostealer script is not novel technical work, but it traditionally required a lot of developer time or buying specialized tools off an underground forum. By using an LLM inside a victim’s network, attackers can generate functional, custom scripts on demand instead of relying on pre-built malware.
While accessing these models still carries some cost, the barrier to entry is lower than ever. We often see average attackers leveraging these tools to work faster and eliminate manual coding tasks. It fundamentally shifts the economics of an attack by making the execution phase highly accessible, allowing lower-skilled actors to scale their operations without needing deep technical expertise.
Huntress reported a 277% year-over-year increase in remote monitoring and management tool abuse. Why have legitimate RMM platforms become such an attractive alternative to custom malware, and what advantages do they give attackers once they gain access?
There is nothing inherently evil about RMM tools. They serve a vital purpose, helping IT admins and MSPs work efficiently and support users. The problem is that custom malware gets flagged by traditional antivirus software, whereas RMMs are signed, trusted, and bypass standard security controls. Even better for an attacker, RMM agents naturally run with elevated admin privileges and provide built-in persistence, meaning they stay active across system reboots without requiring extra effort to hold access.
Once an attacker gets their own RMM agent running on a target machine, they essentially have a remote control console. They can transfer files back and forth seamlessly, and they often set up their RMM instance ahead of time with malicious scripts and tools ready to execute across endpoints with a single click. Because the application itself is legitimate, defenders cannot simply block the tool. It forces security teams to stop relying on trusted software lists and focus on observing behavior, which makes 24/7 monitoring essential if you want to catch an attacker abusing administrative software.
One technique Huntress has documented is attackers “daisy-chaining” multiple RMM tools to fragment telemetry and create redundant access. Can you walk us through how an intrusion like this typically unfolds and why it can be so difficult for defenders to recognize what is happening?
Here is one particularly interesting example that we caught and wrote a blog about. The intrusion started with a phishing email using a fake business document lure, such as a “Network Solutions Agreement”. The victim clicked the link in the email, landed on a page gated by a fake CAPTCHA, and downloaded what they thought was a standard PDF or service document. In reality, the download was an executable installer. When the user double-clicked the file to open the agreement, they silently installed a legitimate but unauthorized remote access tool called Tiflux RMM.
Once the threat actor got that initial Tiflux agent running, they immediately used its built-in capabilities to push down and sideload additional tools like Splashtop, ScreenConnect, and UltraVNC onto the exact same machine. It is important to note that these are all additional legitimate RMM tools themselves.
Attackers do this to help establish super persistence. They know that if an IT admin spots one unauthorized remote tool and uninstalls it, the attacker still has two or three other active backdoors waiting. This tactic fragments the telemetry. Security platforms just see separate, signed administrative applications running simultaneously and phoning home to legitimate vendor infrastructure. Unless a defender is actively monitoring that activity and asking why a single endpoint suddenly needs multiple distinct remote management agents installed in the same afternoon, it just blends right into the noise of routine IT maintenance.
Beyond 24/7 behavioral monitoring, another effective control here is Endpoint Security Posture Management and application control. If an organization explicitly defines which specific flavors of RMMs are authorized to run on their machines, they can restrict unapproved tools from executing at all, stopping the daisy-chain before it starts.
Traditional security models often distinguish between trusted and untrusted software, but RMM attacks exploit tools that may be legitimate, signed, and already approved by IT. Does this mean defenders need to move away from asking whether an application is trusted and focus instead on whether its behavior is trusted?
Yes, absolutely. The traditional model of binary trust is broken. Binary signing was designed to verify that an application came from a legitimate vendor, not to guarantee that its actions are safe.
If an attacker uses a signed RMM binary or a living-off-the-land utility, the executable itself is clean, but the activity behind it is malicious. Defenders have to move past asking “Is this file trusted?” and start focusing on whether its behavior makes sense in context. Having an RMM run on a machine is not inherently evil. But if that tool suddenly starts dropping additional utilities, spawning unfamiliar scripts, or executing at 2 AM under a service account, that is an indicator of compromise. Context and behavior have always mattered, but with attackers relying so heavily on trusted software, they matter now more than ever.
As AI compresses the time between discovering a vulnerability or configuration gap and exploiting it, how does that change the role of security posture management? Which identity, endpoint, and remote-access weaknesses should organizations be prioritizing before attackers have an opportunity to exploit them?
Security posture management matters now more than ever because it helps organizations proactively close the common security gaps attackers exploit every day. It complements detection and response by continuously identifying and addressing issues such as misconfigurations, excessive permissions, unauthorized applications, and other weaknesses across endpoints and identities. When the window between vulnerability disclosure and automated scanning shrinks down to hours, reducing those gaps before they can be exploited becomes essential. Organizations need to maintain a consistent, hardened posture across environments to reduce the number of incidents they face and limit what an attacker can reach if they get through the door.
While almost every cyber vendor will happily “educate” (or scare) the public about whichever threat or security domain they happen to sell solutions for, the actual priority list is something only each individual organization can answer. Every business values different components of the CIA triad, and every company has sensitive data stored or accessed in different places. The real advice most businesses need to hear is to evaluate which data, assets, or production systems are most critical to their operations, and then work backward from there to protect what matters most.
There is growing concern about AI agents eventually carrying out significant portions of the attack lifecycle autonomously. What capabilities would need to improve before autonomous cyberattacks become meaningfully more dangerous than today’s AI-assisted attacks, and are defenders preparing quickly enough?
The concern about autonomous agents implies this is a future problem, but we are already seeing them execute attacks today. However, these autonomous attacks are far from perfect. Look at the recent JadePuffer ransomware campaign. It was run entirely by an AI agent, but it made glaring mistakes. The model hallucinated parts of the operation, including making up a fake Bitcoin wallet address for the ransom note. Even more ridiculous, the AI forgot to save or transmit the encryption key. It just printed the randomly generated key to a temporary console output and closed the session, making it mathematically impossible for the victim to recover their data even if they paid.
The grounded reality is that from a tactics, techniques, and procedures perspective, autonomous AI, automated scripts, and human attackers are all doing the exact same thing. They are still exploiting the same types of internet-exposed, vulnerable, and misconfigured assets. AI still has to go through the standard kill chain. It still needs to achieve initial access, raise permissions, move laterally, and execute its payload. In the JadePuffer case, it just used a known, unpatched vulnerability to get its foot in the door and then stumbled its way through the network.
Because the core mechanics of the attack have not changed, defenders do not need to throw out their current playbook. You still need layered security, basic IT hygiene, and strong behavioral detection. What changes is the pressure and speed at which defenders must process events and triage alerts. When an AI agent makes a mistake, reads the error log, and self-corrects in seconds to keep pushing forward, defenders simply have to adapt to the fast speed that cyber attacks are moving to now.
Attackers are adopting AI, but defenders have access to many of the same technologies. Where do you see AI providing the greatest defensive advantage: detecting unusual behavior, analyzing telemetry, prioritizing vulnerabilities, automating response, or somewhere else entirely?
First, it is highly effective for blue team correlation. Security teams do not have a data problem; they have a noise problem. AI is incredibly useful for pulling the needles out of the haystack, taking disparate, seemingly unrelated events, and stitching them together to tell the full story of an intrusion.
Second, AI is doing for defenders exactly what it does for attackers by bridging the skills gap. When a junior SOC analyst is faced with a heavily obfuscated script or a complex alert, they can use an LLM to instantly translate that data into plain English. It removes the manual grunt work of reverse engineering and allows defenders to triage and respond much faster.
Finally, looking beyond reactive defense, there is a tremendous opportunity for red and purple teams to use AI to proactively test and validate defenses. Fully scoped, human-led penetration tests are still an absolute must. But AI lowers the technical bar for running smaller, atomic tests to validate different layers in a security stack. Instead of waiting for an annual pen-test or crossing your fingers when a real organic threat shows up, organizations can use AI to safely and frequently test their own environments to see what actually breaks.
Looking ahead, do you expect the cybersecurity industry to enter an environment where both attackers and defenders increasingly operate through autonomous agents? If so, what will ultimately determine who has the advantage when machines on both sides can identify and react to threats at machine speed?
AI is definitely shaking up the underground economy. The darknet market for cybercrime-as-a-service has thrived on selling malware and scripts because they traditionally required specialized developer skills to create. AI changes the accessibility and economics of that ecosystem by allowing lower-skilled operators to generate their own tooling on demand. It will likely shut the door on a lot of legacy, off-the-shelf malware vendors. Why would a threat actor pay for generic off-the-shelf malware on an underground forum when they can simply prompt an LLM to generate custom tooling on demand?
When both sides are armed with AI, the advantage comes down to infrastructure, tuning, and oversight. It becomes a battle of who has the compute budget (who has more tokens to burn) and who has the better-tuned model. For threat actors, the test is whether their AI can iteratively troubleshoot, read error logs, and pivot appropriately when an exploit fails. For defenders, the test is whether your detectors are actually tuned and if your data pipelines are properly hooked up to feed the AI the right telemetry in real time. But more importantly, it depends on whether you have human experts with business context in the loop to make the critical judgment calls, such as reading an AI’s interpretation of an incident and deciding whether to isolate a compromised machine or take a production system offline to contain a threat.
Ultimately, AI does not change the fundamentals of security. The businesses that will fare the best are the ones that continue to get the little things right. It comes down to reducing your attack surface, patching vulnerabilities, hardening configurations, monitoring 24/7, assuming compromise, and proactively testing your defenses regularly. AI is just an accelerator. Whoever executes the basics faster and more consistently is the one who wins.
Thank you for the great interview, readers who wish to learn more should visit Huntress.
