Two new reports from Kaspersky and CrowdStrike point to a concerning shift in how hackers are using AI: agents are now helping carry out real-world intrusions, from navigating compromised networks to harvesting credentials and escalating privileges. Which of course is bad.
Kevin Surace, CEO, TokenCore
“What CrowdStrike is documenting is an important turning point. AI is no longer simply helping attackers write code or craft better phishing emails. It is beginning to operate inside compromised environments, make decisions, move laterally, harvest credentials and escalate privileges on its own.
“That makes identity compromise dramatically more dangerous.
“In one of the cases CrowdStrike studied, an AI agent moved laterally to a domain controller and harvested credentials in roughly 20 minutes without additional human direction. In another, the AI discovered exposed credentials, obtained an identity and access management token, escalated to cluster administrator privileges and then extracted additional database credentials.
This changes the economics and speed of cybercrime. A stolen identity used to give one hacker a foothold. Increasingly, that same identity can give an autonomous agent a foothold, and that agent can operate continuously, explore the environment, find additional identities and privileges and determine its next move at machine speed.
“The report makes another point that should concern every CISO. CrowdStrike found that engineering sophistication was not necessarily the limiting factor. Even relatively low skill attackers were able to conduct successful intrusions using agentic AI. CrowdStrike concluded that token budget, rather than attacker skill, may increasingly become the constraint on scale.
So the security question is changing. It is no longer enough to ask whether someone has the correct password, MFA approval or passkey attack. Organizations have to know whether the person behind a privileged action is actually the authorized human.
“In an agentic world, identity has to become a hard security boundary. For critical access, privilege escalation and consequential actions, enterprises will increasingly need cryptographic authentication tied to the legitimate application and a physical biometric from the authorized person.
“AI is making sophisticated cyber operations faster, cheaper and accessible to far more attackers. That means a compromised identity is becoming more valuable than ever. If attackers can automate nearly everything that happens after they get inside, then preventing them from becoming a trusted user in the first place becomes one of the most important defenses we have.”
John Strand, Owner, Black Hills Information Security ()
“I believe what we’re seeing is the new normal. This is the evolution of cyberattacks, and organizations need to stop treating it as an aberration.
“For years, security has leaned on the idea that you don’t have to be the fastest person running from the bear. You just have to be faster than the slowest person. Well, now the bear has a jetpack and an AK-47, and it attacks indiscriminately.
“When organizations start treating swarms of AI agents and their relentless persistence as the new normal, it should fundamentally change how they defend themselves.
“I’ve been telling every organization I talk to the same thing. Invest in cyber deception. Put tripwires in place. Then reduce your attack surface as much as possible. If an externally exposed service isn’t required, firewall it off or disconnect it. That applies even if it has no known critical vulnerabilities.
“Don’t give attackers any more room to get into your organization than you absolutely have to.”
Jacob Krell, Sr. Director: Secure AI Solutions & Cybersecurity, Suzu Labs
“Kaspersky caught attackers running CyberStrikeAI, an open-source framework where a large language model (LLM) on the command-and-control (C2) server writes the commands while the implant on the victim’s machine just runs them. Education targets in Latin America and Africa went from a compromised SQL Server to domain administrator in three to four days, with final exfiltration under an hour. CrowdStrike separately identified four new threat actors using AI agents, including one that harvested domain controller credentials in about twenty minutes with no human involved.
“The architecture is what makes this productizable. All the intelligence lives on the attacker’s server, so the implant holds nothing worth seizing. Ransomware-as-a-Service (RaaS) proved buyers with crypto and no skills will rent an intrusion stack, and agentic frameworks remove the last labor input, the operator who knows which command comes next.
“My prediction is the pairing. Darknet listings will bundle a CyberStrikeAI-style harness with an uncensored model running on the seller’s hardware, sold for cryptocurrency. The customer brings a target. The seller brings the model that decides the commands. That is the hacking-as-a-service evolution I think we will begin to see.”
Darin Fredde, Sr. Director of Technical Marketing Engineering, Ridge Security
“What strikes me about these reports isn’t that AI has invented some revolutionary new form of hacking. Much of the underlying tradecraft is familiar: reconnaissance, credential access, privilege escalation, lateral movement, and data collection. We’ve been testing these techniques for decades.
“What’s changing is how attackers orchestrate them. AI is moving beyond helping write phishing emails or generate scripts toward interpreting results, selecting tools, and coordinating subsequent actions. That can reduce the hands-on effort required to execute familiar attack workflows, although we should be careful about claiming specific speed improvements without comparable human-led testing.
“Having spent time in penetration testing, full-scope adversary emulation, and security operations roles, I see an equally important question for defenders: If attackers are using AI to accelerate traditional tradecraft, why aren’t we applying those same capabilities to continuously test, challenge, and improve our defenses?
“The opportunity isn’t simply to automate penetration testing. It’s to bring offensive testing and defensive operations closer together through continuous purple teaming: emulate realistic attack paths, observe what our controls detect, identify the gaps, remediate them, and independently verify the results.
“Of course, there’s a certain irony in attackers adopting automation while defenders are still scheduling the next quarterly security assessment. The attackers apparently didn’t get the calendar invitation.
“AI doesn’t eliminate the need for experienced security professionals. It amplifies what they can accomplish, provided we maintain appropriate authorization, operational boundaries, and independent verification.
“The real advantage won’t be long simply to whoever automates attacks fastest. It will belong to those who can turn that same speed into measurable improvements in detection, containment, and remediation.”
The question is not if you should defend against AI. It’s when. Because if you don’t you will get pwned. It is that simple.
Related
This entry was posted on November 3, 2026 at 7:24 am and is filed under Commentary with tags AI. You can follow any responses to this entry through the RSS 2.0 feed.
You can leave a response, or trackback from your own site.
AI is changing how hackers operate. Are defenders keeping up?
Two new reports from Kaspersky and CrowdStrike point to a concerning shift in how hackers are using AI: agents are now helping carry out real-world intrusions, from navigating compromised networks to harvesting credentials and escalating privileges. Which of course is bad.
Kevin Surace, CEO, TokenCore
“What CrowdStrike is documenting is an important turning point. AI is no longer simply helping attackers write code or craft better phishing emails. It is beginning to operate inside compromised environments, make decisions, move laterally, harvest credentials and escalate privileges on its own.
“That makes identity compromise dramatically more dangerous.
“In one of the cases CrowdStrike studied, an AI agent moved laterally to a domain controller and harvested credentials in roughly 20 minutes without additional human direction. In another, the AI discovered exposed credentials, obtained an identity and access management token, escalated to cluster administrator privileges and then extracted additional database credentials.
This changes the economics and speed of cybercrime. A stolen identity used to give one hacker a foothold. Increasingly, that same identity can give an autonomous agent a foothold, and that agent can operate continuously, explore the environment, find additional identities and privileges and determine its next move at machine speed.
“The report makes another point that should concern every CISO. CrowdStrike found that engineering sophistication was not necessarily the limiting factor. Even relatively low skill attackers were able to conduct successful intrusions using agentic AI. CrowdStrike concluded that token budget, rather than attacker skill, may increasingly become the constraint on scale.
So the security question is changing. It is no longer enough to ask whether someone has the correct password, MFA approval or passkey attack. Organizations have to know whether the person behind a privileged action is actually the authorized human.
“In an agentic world, identity has to become a hard security boundary. For critical access, privilege escalation and consequential actions, enterprises will increasingly need cryptographic authentication tied to the legitimate application and a physical biometric from the authorized person.
“AI is making sophisticated cyber operations faster, cheaper and accessible to far more attackers. That means a compromised identity is becoming more valuable than ever. If attackers can automate nearly everything that happens after they get inside, then preventing them from becoming a trusted user in the first place becomes one of the most important defenses we have.”
John Strand, Owner, Black Hills Information Security ()
“I believe what we’re seeing is the new normal. This is the evolution of cyberattacks, and organizations need to stop treating it as an aberration.
“For years, security has leaned on the idea that you don’t have to be the fastest person running from the bear. You just have to be faster than the slowest person. Well, now the bear has a jetpack and an AK-47, and it attacks indiscriminately.
“When organizations start treating swarms of AI agents and their relentless persistence as the new normal, it should fundamentally change how they defend themselves.
“I’ve been telling every organization I talk to the same thing. Invest in cyber deception. Put tripwires in place. Then reduce your attack surface as much as possible. If an externally exposed service isn’t required, firewall it off or disconnect it. That applies even if it has no known critical vulnerabilities.
“Don’t give attackers any more room to get into your organization than you absolutely have to.”
Jacob Krell, Sr. Director: Secure AI Solutions & Cybersecurity, Suzu Labs
“Kaspersky caught attackers running CyberStrikeAI, an open-source framework where a large language model (LLM) on the command-and-control (C2) server writes the commands while the implant on the victim’s machine just runs them. Education targets in Latin America and Africa went from a compromised SQL Server to domain administrator in three to four days, with final exfiltration under an hour. CrowdStrike separately identified four new threat actors using AI agents, including one that harvested domain controller credentials in about twenty minutes with no human involved.
“The architecture is what makes this productizable. All the intelligence lives on the attacker’s server, so the implant holds nothing worth seizing. Ransomware-as-a-Service (RaaS) proved buyers with crypto and no skills will rent an intrusion stack, and agentic frameworks remove the last labor input, the operator who knows which command comes next.
“My prediction is the pairing. Darknet listings will bundle a CyberStrikeAI-style harness with an uncensored model running on the seller’s hardware, sold for cryptocurrency. The customer brings a target. The seller brings the model that decides the commands. That is the hacking-as-a-service evolution I think we will begin to see.”
Darin Fredde, Sr. Director of Technical Marketing Engineering, Ridge Security
“What strikes me about these reports isn’t that AI has invented some revolutionary new form of hacking. Much of the underlying tradecraft is familiar: reconnaissance, credential access, privilege escalation, lateral movement, and data collection. We’ve been testing these techniques for decades.
“What’s changing is how attackers orchestrate them. AI is moving beyond helping write phishing emails or generate scripts toward interpreting results, selecting tools, and coordinating subsequent actions. That can reduce the hands-on effort required to execute familiar attack workflows, although we should be careful about claiming specific speed improvements without comparable human-led testing.
“Having spent time in penetration testing, full-scope adversary emulation, and security operations roles, I see an equally important question for defenders: If attackers are using AI to accelerate traditional tradecraft, why aren’t we applying those same capabilities to continuously test, challenge, and improve our defenses?
“The opportunity isn’t simply to automate penetration testing. It’s to bring offensive testing and defensive operations closer together through continuous purple teaming: emulate realistic attack paths, observe what our controls detect, identify the gaps, remediate them, and independently verify the results.
“Of course, there’s a certain irony in attackers adopting automation while defenders are still scheduling the next quarterly security assessment. The attackers apparently didn’t get the calendar invitation.
“AI doesn’t eliminate the need for experienced security professionals. It amplifies what they can accomplish, provided we maintain appropriate authorization, operational boundaries, and independent verification.
“The real advantage won’t be long simply to whoever automates attacks fastest. It will belong to those who can turn that same speed into measurable improvements in detection, containment, and remediation.”
The question is not if you should defend against AI. It’s when. Because if you don’t you will get pwned. It is that simple.
Share this:
Like this:
Related
This entry was posted on November 3, 2026 at 7:24 am and is filed under Commentary with tags AI. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.